Construction method and system of first-aid medical knowledge base based on medical knowledge graph

Abstract

The application discloses a first aid medical knowledge base construction method and system based on a medical knowledge graph, and comprises the following steps: collecting emergency first aid clinical guideline data, expert consensus data, reference data and first aid diagnosis and treatment data, performing cleaning, alignment and standardization processing, and generating a first aid medical knowledge source data sequence; based on the first aid medical knowledge source data sequence, constructing a first aid diagnosis and treatment event development graph, and generating a time slice graph sequence; constructing a double evolution EvolveGCN model; based on the time slice graph sequence and the first aid diagnosis and treatment event development graph, generating event development representation, a stable knowledge subgraph sequence and a state knowledge subgraph sequence; based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, generating a joint relationship representation; based on the joint relationship representation, generating a correction relationship set; and based on the correction relationship set, generating a first aid medical knowledge base. The application improves the knowledge storage accuracy and time sequence consistency.

Description

Technical Field

[0001] This invention relates to the field of medical informatics, and in particular to a method and system for constructing an emergency medical knowledge base based on medical knowledge graphs. Background Technology

[0002] In emergency situations, patients' conditions change rapidly, and timely decision-making is crucial. Relevant information is distributed across emergency clinical guidelines, expert consensus, references, and emergency treatment records. Current emergency medical information processing methods primarily rely on reviewing medical records, analyzing single test results, and using localized rules for retrieval. They lack unified organization, temporal correlation, and structured organization of multi-source medical information, resulting in unclear correspondences between treatment subjects, clinical events, and evidence status, hindering the formation of a continuously updated knowledge base.

[0003] Existing methods for constructing medical knowledge bases often focus on static medical knowledge extraction and conventional graph modeling, paying insufficient attention to the characteristics of knowledge relationships changing over time during emergency care. They struggle to simultaneously address the evolving relationships between stable medical knowledge and dynamic diagnostic and treatment knowledge. Furthermore, existing technologies generally lack effective unified verification mechanisms for supporting relationships, conflicting relationships, source differences, and time-lapse issues among multi-source information. This can easily lead to distorted knowledge associations, residual expired knowledge, and inconsistencies between the data and the final input, thus affecting the accuracy, completeness, and usability of the emergency medical knowledge base.

[0004] Therefore, how to provide a method and system for constructing an emergency medical knowledge base based on medical knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for constructing an emergency medical knowledge base based on medical knowledge graphs. This invention constructs an emergency medical event unfolding graph and a time slice graph sequence based on emergency clinical guidelines, expert consensus data, reference data, and emergency treatment data. It then constructs a dual-evolutionary EvolveGCN model and performs event unfolding encoding, dual-branch relationship evolution, and support conflict correction processing on relational connections through event unfolding encoding units, stable knowledge evolution units, state knowledge evolution units, and support conflict correction units. This generates a corrected relation set, and further, based on the corrected relation set, it performs entry condition judgments to generate an emergency medical knowledge base. This method has the advantages of clear relational expression, accurate time evolution characterization, strong relation correction capability, and high-quality knowledge entry.

[0006] The method for constructing an emergency medical knowledge base based on a medical knowledge graph according to an embodiment of the present invention includes the following steps: Collect emergency and critical care clinical guidelines data, expert consensus data, reference data, and emergency treatment data; perform cleaning, alignment, and standardization processing to generate an emergency medical knowledge source data sequence. Based on the emergency medical knowledge source data sequence, an emergency diagnosis and treatment event unfolding graph is constructed, and a time slice graph sequence is generated based on the time distribution of each node and each relationship connection. A dual-evolution EvolveGCN model is constructed, including an event unfolding encoding unit, a stable knowledge evolution unit, a state knowledge evolution unit, and a support conflict correction unit. Based on the time slice graph sequence, the emergency medical event unfolding graph is input into the event unfolding encoding unit. The node connection structure and relation connection structure in each time slice are encoded to generate the event unfolding representation. The relation connections reflecting long-term medical relationships and the relation connections reflecting changes in the emergency process are separated to generate a stable knowledge subgraph sequence and a state knowledge subgraph sequence. Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the first parameter evolution process and the second parameter evolution process are performed in the stable knowledge evolution unit and the state knowledge evolution unit, respectively, to generate a joint relation representation; Based on the joint relation representation, a support relation graph and a conflict relation graph are constructed, and the support conflict correction unit is input to perform dual graph propagation processing and reverse correction processing to generate a set of correction relations; Based on the set of correction relationships, the entry conditions are judged to generate an emergency medical knowledge base.

[0007] Optionally, the generation of emergency medical knowledge source data sequences includes: Data from emergency and critical care clinical guidelines, expert consensus, references, and emergency treatment data were collected. The emergency treatment data included chief complaint data, vital signs data, laboratory data, imaging data, diagnostic data, treatment data, and time-stamped data. Following unified field rules, field splitting was performed on emergency and critical care clinical guideline data, expert consensus data, reference data, and emergency treatment data to generate a basic record set. Based on the unified time indexing rules, time alignment processing is performed on the basic record set to generate a time-aligned record set; Based on the unified medical terminology mapping rules, terminology standardization processing is performed on the time-aligned record set to generate a standardized record set; Based on a standardized record set, according to the data ownership relationship in the same continuous emergency process, the same medical object is assigned an identifier, written into the corresponding data record, and an object-marked record set is generated. Based on the object-tagged record set, each data is mapped to a corresponding clinical event, each clinical event is assigned a clinical event tag, written to the corresponding data record, and an event-tagged record set is generated. Based on the event marker record set, merge processing is performed within the same time slice to generate evidence status markers, write the corresponding data records, and generate an evidence marker record set; Based on the evidence-tagged record set, a structured organization is performed to generate a sequence of emergency medical knowledge source data.

[0008] Optionally, the generation of the time slice sequence includes: Based on the emergency medical knowledge source data sequence, the fields of diagnosis and treatment object identifier, clinical event marker, evidence status marker and source marker are read one by one to establish diagnosis and treatment object node, clinical event node and evidence status node; Write the treatment object identifier and time index field into the treatment object node, write the clinical event marker, time index field and source marker field into the clinical event node, and write the evidence status marker, time index field and source marker field into the evidence status node to form the node attribute fields of the treatment object node, clinical event node and evidence status node; Establish the chronological order of occurrence based on the clinical event nodes corresponding to the same patient identification; Establish evidence support relationships based on the evidence status nodes and clinical event nodes corresponding to the same diagnosis and treatment object identifier; Establish a diagnosis and treatment trigger relationship based on the diagnosis and treatment object node and clinical event node corresponding to the same diagnosis and treatment object identifier; Based on the diagnosis and treatment object node, clinical event node, evidence status node, sequence of occurrence, evidence support relationship and diagnosis and treatment trigger relationship, construct an emergency diagnosis and treatment event unfolding diagram; Based on the time distribution of clinical event nodes in the emergency medical event unfolding diagram, time slices are divided according to the occurrence point of the clinical event node. Based on the temporal order between adjacent time slices, the diagnosis and treatment object nodes in the previous time slice are inherited into the next time slice. The clinical event nodes and evidence status nodes in the corresponding time slices of the next time slice are retained in the next time slice. Based on the inherited diagnosis and treatment object nodes, the clinical event nodes and evidence status nodes in the next time slice, the corresponding chronological order, evidence support, and diagnosis and treatment triggering relationships are re-established to generate a time slice sequence.

[0009] Optionally, constructing the dual-evolution EvolveGCN model includes: The basic structure of the original EvolveGCN model is invoked, including a sequence of graph convolutional layers that process graph sequences in time order and a sequence of parameter evolution layers connected to the graph convolutional layer sequence. Based on the emergency medical event unfolding diagram and time slice sequence, the basic structure of the original EvolveGCN model is rewritten at the input end, and an event unfolding encoding unit is added. Based on the graph convolutional layer sequence and parameter evolution layer sequence, a two-branch rewriting is performed to split the single-path time evolution structure into a stable knowledge evolution unit and a state knowledge evolution unit. Based on the stable knowledge evolution unit and the state knowledge evolution unit, the basic structure of the original EvolveGCN model is extended at the output end, and a support conflict correction unit is added. Based on the event unfolding encoding unit, stable knowledge evolution unit, state knowledge evolution unit, support conflict correction unit, and the connection relationships between each unit, a dual-evolution EvolveGCN model is generated.

[0010] Optionally, the generation of stable knowledge subgraph sequences and state knowledge subgraph sequences includes: Based on the time slice graph sequence, the diagnosis and treatment object nodes, clinical event nodes, evidence status nodes and the corresponding relationship connections of each node are extracted from each time slice and combined to form the diagnosis and treatment object node input vector, clinical event node input vector and evidence status node input vector. According to the chronological order of the time slices, the input vectors of the diagnosis and treatment object nodes, the input vectors of the clinical event nodes, the input vectors of the evidence state nodes, and the relational connections are written into the event expansion encoding unit, and the encoding process is performed to obtain the node representation and relational representation corresponding to each time slice; Based on the node representations and relationship representations in each time slice, aggregation processing is performed to combine and generate the time slice representation corresponding to the current time slice, and then arranged in chronological order to generate the event expansion representation; Based on the event unfolding representation, relation separation processing is performed on the relational connections in each time slice to determine the relational connections that reflect long-term medical relationships and the relational connections that reflect changes in the emergency process. Based on relational connections reflecting long-term medical relationships and corresponding time slice nodes of the diagnosis and treatment object, clinical event, and evidence status, a stable knowledge subgraph sequence is constructed. Based on relational connections reflecting changes in the emergency treatment process and corresponding time slice nodes of the diagnosis and treatment object, clinical event, and evidence status, a state knowledge subgraph sequence is constructed.

[0011] Optionally, the generation of the union representation includes: Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the stable knowledge subgraph corresponding to each time slice is input into the stable knowledge evolution unit, and the state knowledge subgraph corresponding to each time slice is input into the state knowledge evolution unit. In the stable knowledge evolution unit, the first parameter evolution processing is performed on the current input according to the time slice order to generate a stable relation representation corresponding to the current time slice and extract the medical relation boundary; In the state knowledge evolution unit, the second parameter evolution process is performed on the current input in the order of time slices to generate a time-varying relation representation; Within each time slice, the time-varying relation representation corresponding to the current time slice is introduced into the relation connection parameter update stage of the first parameter evolution process to generate a stable relation representation after cross-introduction of the time-varying relation representation; Within each time slice, the medical relationship boundary corresponding to the current time slice is introduced into the relationship connection parameter update stage of the second parameter evolution process to generate a time-varying relationship representation after the cross-introduction of the medical relationship boundary; In the stable knowledge evolution unit and the state knowledge evolution unit, the stable relation representation after the cross-introduction of time-varying relation representation and the time-varying relation representation after the cross-introduction of medical relation boundary are aligned and combined to generate a joint relation representation.

[0012] Optionally, the set of generated correction relationships includes: Based on the union relation representation, the relation connection, relation category, connection object, time index field, source tag field and evidence status tag within each time slice are extracted, and the union relation representation corresponding to each relation connection is determined as the relation representation of the corresponding relation connection; Within the same time slice, support determination is performed on each pair of relational connections to construct connection pairs in the support relation graph, and the corresponding relation representation is configured on the connection pair to generate the support relation graph. Within the same time slice, conflict determination is performed on each pair of relational connections to construct connection pairs in the conflict relation graph, and the corresponding relation representation is configured on the connection pair to generate the conflict relation graph. Input the support relationship diagram and the conflict relationship diagram into the support conflict correction unit. Perform support propagation processing on the relationship representation in the support relationship diagram according to the connection direction within the diagram, and perform conflict propagation processing on the relationship representation in the conflict relationship diagram according to the connection direction within the diagram, generating support propagation results and conflict propagation results respectively. Based on the support propagation results and conflict propagation results, a dual-graph merging update is performed on each relation connection within the same time slice to generate an initial corrected relation representation; Based on the consistent connection objects and relationship categories between adjacent time slices, reverse correction processing is performed on the initial correction relationship representation in the later time slice to generate a cross-time slice correction relationship representation; Based on the time slice order, the cross-time slice correction relationship representations in each time slice are summarized to generate a correction relationship set.

[0013] Optionally, the generation of the emergency medical knowledge base includes: Based on the set of corrected relations, extract the time index field, relation category, connection object, source tag field, correction result and joint relation representation corresponding to each relation connection, and generate an entry judgment sequence; Perform time consistency checks on each relation connection in the inbound decision sequence; Perform a relationship consistency check on each relation connection in the inbound decision sequence; Perform source consistency checks on each relation connection in the inbound decision sequence; For current relational connections that meet the time consistency, relational consistency, and source consistency requirements, they are determined to meet the entry conditions. A formal relational record area is constructed, and the current relational connection, the corresponding connection object, relational category, time index field, source tag field, and correction result are written into the formal relational record area to generate an entry relational connection. For current relational connections that do not meet the requirements of time consistency, relational consistency, and source consistency, perform verification marking processing, construct a relational record area to be verified, and write the current relational connection along with the corresponding connection object, relational category, time index field, source mark field, correction result and failed consistency judgment item into the relational record area to be verified, thus generating a relational connection to be verified; An emergency medical knowledge base is generated based on the formal relationship record area and the relationship record area to be verified.

[0014] The emergency medical knowledge base construction system based on medical knowledge graph according to an embodiment of the present invention is characterized by comprising: The emergency medical knowledge source data processing module is used to collect emergency and first aid clinical guidelines data, expert consensus data, reference data and emergency diagnosis and treatment data, perform cleaning, alignment, standardization processing and structured organization, and generate emergency medical knowledge source data sequences. The time slice graph construction module is used to establish treatment object nodes, clinical event nodes, and evidence status nodes based on the emergency medical knowledge source data sequence, construct an emergency treatment event unfolding graph, and generate a time slice graph sequence; The dual-evolution EvolveGCN model building module is used to call the basic structure of the original EvolveGCN model, perform structural rewriting, and generate the dual-evolution EvolveGCN model. The event unfolding separation module is used to perform encoding processing in the event unfolding encoding unit based on the time slice graph sequence, generate the event unfolding representation, and separate the generated stable knowledge subgraph sequence and state knowledge subgraph sequence. The joint relation evolution module is used to generate joint relation representations by performing first parameter evolution processing and second parameter evolution processing on stable knowledge subgraph sequences and state knowledge subgraph sequences, respectively. The support conflict correction module is used to construct support relationship graphs and conflict relationship graphs based on the joint relationship representation, and to perform dual-graph propagation processing and reverse correction processing to generate a set of corrected relationships; The inbound judgment module is used to perform inbound condition judgment based on the set of correction relationships, and generate inbound relationship connections and relationship connections to be verified respectively. The emergency medical knowledge base generation module is used to generate an emergency medical knowledge base based on the relationship record content in the formal relationship record area and the relationship record area to be verified.

[0015] The beneficial effects of this invention are: First, this invention constructs an emergency medical event unfolding diagram and a time-slice diagram sequence based on emergency medical knowledge source data sequences. It connects treatment object nodes, clinical event nodes, evidence status nodes, and their corresponding relationships into a unified temporal expression framework, enabling the continuous organization of the evolutionary relationships in the emergency process according to time sequence. Compared to existing technologies that only construct static associations for single treatment data, this invention can more accurately reflect the association status of emergency clinical guidelines data, expert consensus data, reference data, and emergency medical data within the same emergency process, improving the structured expression and temporal association expression capabilities of the emergency medical knowledge base for the emergency process.

[0016] Secondly, this invention constructs a dual-evolution EvolveGCN model. Through the coordinated operation of event unfolding encoding units, stable knowledge evolution units, state knowledge evolution units, and supporting conflict correction units, it performs event unfolding encoding, bi-branch parameter evolution, and joint relation generation processing on stable knowledge subgraph sequences and state knowledge subgraph sequences. Furthermore, a linkage update mechanism is established between the first and second parameter evolution processes, enabling the simultaneous characterization of long-term medical relationships and changing relationships in emergency care within the same technical framework. Based on this processing method, the joint relation representation more fully characterizes the stable boundaries and time-varying features of medical relationships, which is beneficial for improving the accuracy of relationship connection identification, the continuity of emergency care process characterization, and the reliability of complex diagnostic and treatment relationship modeling.

[0017] Furthermore, this invention constructs supporting and conflicting relationship graphs based on joint relationship representation, and generates a corrected relationship set through dual-graph propagation processing and reverse correction processing. This set is then combined with the formal relationship record area and the relationship record area to be verified to generate an emergency medical knowledge base. This allows for the classification, correction, and partitioning of mutually validating, contradictory, time-expired, and inconsistently sourced relationship connections before they are added to the database. Based on this processing method, the probability of erroneous relationships entering the emergency medical knowledge base is reduced, improving the credibility, consistency, and traceability of relationship connections within the knowledge base, and providing a data foundation for the subsequent verification and continuous updating of relationship connections to be verified. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is an overall flowchart of the emergency medical knowledge base construction method and system based on medical knowledge graph proposed in this invention; Figure 2 This is a schematic diagram of the structural connections of the dual-evolution EvolveGCN model proposed in this invention; Figure 3 This is a schematic diagram of the joint correction process of the support relationship diagram and conflict relationship diagram proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 The method for constructing an emergency medical knowledge base based on medical knowledge graphs includes the following steps: Data from emergency and critical care clinical guidelines, expert consensus, references, and emergency treatment were collected. The emergency treatment data included chief complaint data, vital sign data, laboratory test data, imaging data, diagnostic data, treatment data, and time-stamped data. Based on unified field rules, unified time index rules, and unified medical terminology mapping rules, the emergency and critical care clinical guidelines, expert consensus, references, and emergency treatment data were cleaned, aligned, and standardized. From the clinical guidelines, expert consensus, and references, descriptions of the patient's condition, clinical events, evidence status, time index fields, and source marker fields were extracted. From the emergency treatment data, chief complaint data, vital sign data, laboratory test data, imaging data, diagnostic data, treatment data, and time-stamped data were extracted. These data were then combined with patient identification, clinical event markers, and evidence status markers for unified structured organization to generate a sequence of emergency medical knowledge source data. Based on the emergency medical knowledge source data sequence, the diagnosis and treatment objects, clinical events, and evidence status corresponding to each time point are decomposed into diagnosis and treatment object nodes, clinical event nodes, and evidence status nodes. An emergency diagnosis and treatment event unfolding diagram is constructed according to the order of occurrence, evidence support relationship, and diagnosis and treatment trigger relationship. Based on the time distribution of the connection between each node and each relationship in the emergency diagnosis and treatment event unfolding diagram, a time slice map sequence is generated. A dual-evolution EvolveGCN model is constructed, which includes an event unfolding encoding unit, a stable knowledge evolution unit, a state knowledge evolution unit, and a support conflict correction unit. Based on the time-slice graph sequence, the treatment object nodes, clinical event nodes, and evidence state nodes in the emergency treatment event unfolding graph are input into the event unfolding encoding unit in the dual-evolution EvolveGCN model. The node connection structure and relation connection structure in each time slice are encoded according to the time slice order to generate the event unfolding representation. Based on the event unfolding representation, the relation connections reflecting long-term medical relationships and the relation connections reflecting changes in the emergency treatment process are separated to generate a stable knowledge subgraph sequence and a state knowledge subgraph sequence. Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the first parameter evolution process and the second parameter evolution process are performed in the stable knowledge evolution unit and the state knowledge evolution unit, respectively. The time-varying relation representation obtained by the second parameter evolution process is introduced into the first parameter evolution process, and the medical relation boundary obtained by the first parameter evolution process is introduced into the second parameter evolution process to generate the joint relation representation. Based on the joint relation representation, mutually verified relation connections are constructed into a support relation graph, and contradictory, time-invalid, and inconsistent relation connections are constructed into a conflict relation graph. The support relation graph and the conflict relation graph are input into the support conflict correction unit in the dual-evolution EvolveGCN model. Dual graph propagation processing is performed within the same time slice, and reverse correction processing is performed between adjacent time slices to generate a set of correction relations. Based on the set of corrected relations, the relation connections in the set are judged according to time consistency, relation consistency, and source consistency. Relation connections that meet the entry conditions are written into the emergency medical knowledge base, and relation connections that do not meet the entry conditions are marked as relation connections to be verified, thus generating the emergency medical knowledge base.

[0021] In this embodiment, generating the emergency medical knowledge source data sequence includes: Data from emergency and critical care clinical guidelines, expert consensus, references, and emergency treatment data were collected. The emergency treatment data included chief complaint data, vital signs data, laboratory data, imaging data, diagnostic data, treatment data, and time-stamped data. Following unified field rules, field splitting was performed on emergency and critical care clinical guideline data, expert consensus data, reference data, and emergency treatment data. Emergency and critical care clinical guideline data, expert consensus data, and reference data were split into descriptions of the treatment object, descriptions of clinical events, descriptions of evidence status, time index fields, and source marker fields. Emergency treatment data were split into chief complaint data, vital signs data, laboratory data, imaging data, diagnostic data, treatment data, and time-marked data. Time index fields and source marker fields were configured in the corresponding data records to generate a basic record set. Based on unified field rules, the basic record set is processed to complete missing fields, remove duplicate fields, and correct abnormal fields, generating a cleaned record set. Based on a unified time indexing rule, time alignment processing is performed on the descriptions of medical subjects, clinical events, evidence status, chief complaints, physical signs, laboratory data, imaging data, diagnostic data, and treatment data in the cleaned record set. Data records whose time stamp data falls within the same time index interval are grouped into the same time slice, and data records that span different time index intervals are split according to the time stamp data and grouped into the corresponding time slices, thus generating a time-aligned record set. Based on the unified medical terminology mapping rules, the descriptions of diagnosis and treatment objects, clinical events, evidence status, chief complaints, signs, laboratory tests, imaging data, diagnostic data, and treatment data in the time-aligned record set are standardized with terms, mapping textual and coded expressions to unified medical terms and generating a standardized record set. Based on a standardized record set, according to the data attribution relationship in the same continuous emergency process, the data records corresponding to the description content of the treatment object, the description content of clinical events, the description content of evidence status, chief complaint data, vital sign data, test data, imaging data, diagnostic data, and treatment data are assigned the same treatment object identifier, and an object-tagged record set is generated. Based on the object-tagged record set, the description content of the diagnosis and treatment object, the description content of the clinical event, the chief complaint data, the sign data, the test data, the imaging data, the diagnostic data, and the treatment data are mapped to the corresponding clinical events. Each clinical event is assigned a clinical event tag according to the time index field, and an event tag record set is generated. Based on the event-marked record set, the evidence status description content, chief complaint data, vital sign data, test data, imaging data, diagnostic data, and treatment data are merged within the same time slice according to the time index field. Based on the merging results, the corresponding data records are assigned evidence status marks, and an evidence-marked record set is generated. Based on the evidence-marked record set, the fields of diagnosis and treatment object identification, clinical event marking, evidence status marking and source marking are structured according to the time index field to generate emergency medical knowledge source data sequence.

[0022] In this embodiment, generating the time slice sequence includes: Based on the emergency medical knowledge source data sequence, the fields of diagnosis and treatment object identification, clinical event marker, evidence status marker and source marker are decomposed according to the time index field. Diagnosis and treatment object nodes are established according to the diagnosis and treatment object identification, clinical event nodes are established according to the clinical event marker, and evidence status nodes are established according to the evidence status marker. Based on the diagnosis and treatment object node, clinical event node, and evidence status node, configure their respective corresponding diagnosis and treatment object identifier, clinical event marker, evidence status marker, time index field, and source marker field to form the node attribute fields of the diagnosis and treatment object node, clinical event node, and evidence status node; Based on the clinical event nodes corresponding to the same patient identification, a chronological relationship is established between the previous and subsequent clinical event nodes according to the order of the time index fields. Based on the evidence status node and clinical event node corresponding to the same diagnosis and treatment object identifier, connect the evidence status node with the same time index field and the source mark field corresponding to the current clinical event node to the current clinical event node to establish evidence support relationship; Based on the diagnosis and treatment object node and clinical event node corresponding to the same diagnosis and treatment object identifier, the earliest clinical event node in the time index field is connected to the diagnosis and treatment object node to establish the first diagnosis and treatment trigger relationship, and subsequent diagnosis and treatment trigger relationships are established between adjacent clinical event nodes based on the order of occurrence. Based on the diagnosis and treatment object node, clinical event node, evidence status node, sequence of occurrence, evidence support relationship and diagnosis and treatment trigger relationship, construct an emergency diagnosis and treatment event unfolding diagram; Based on the time distribution of each node and relationship in the emergency medical event unfolding diagram, the emergency medical event unfolding diagram is divided into time slices according to the time index field. The medical object nodes, clinical event nodes, evidence status nodes, chronological relationships, evidence support relationships and medical triggering relationships within the same time index interval are organized into corresponding time slice diagrams. Based on the temporal order of each time slice, the diagnosis and treatment object nodes in adjacent time slices are continuously preserved. The clinical event nodes, evidence status nodes, chronological relationships, evidence support relationships and diagnosis and treatment trigger relationships in the corresponding time index intervals of each time slice are sequentially organized to generate a time slice sequence.

[0023] In this embodiment, constructing the dual-evolution EvolveGCN model includes: The basic structure of the original EvolveGCN model is invoked. The basic structure of the original EvolveGCN model includes a graph convolutional layer sequence that processes the graph sequence in time order and a parameter evolution layer sequence connected to the graph convolutional layer sequence. The parameter evolution layer sequence is used to recursively update the layer parameters of each graph convolutional layer in the graph convolutional layer sequence at each time slice. Based on the unfolded graph of emergency medical events and the sequence of time slice graphs, the basic structure of the original EvolveGCN model is rewritten at the input end. An event unfolding encoding unit is added before the graph convolutional layer sequence. The node connection structure and relation connection structure in the time slice graph sequence are written into the event unfolding encoding unit, and the output end of the event unfolding encoding unit is connected to the stable knowledge evolution unit and the state knowledge evolution unit, respectively. Based on the graph convolutional layer sequence and parameter evolution layer sequence of the original EvolveGCN model, a two-branch rewrite is performed on the basic structure of the original EvolveGCN model. The single-path time evolution structure in the original EvolveGCN model is split into stable knowledge evolution units and state knowledge evolution units. The stable knowledge evolution unit includes a first graph convolutional layer sequence and a first parameter evolution layer sequence connected to the first graph convolutional layer sequence. The state knowledge evolution unit includes a second graph convolutional layer sequence and a second parameter evolution layer sequence connected to the second graph convolutional layer sequence. The output of the event unpacking encoding unit is simultaneously written into the first graph convolutional layer sequence and the second graph convolutional layer sequence. The layer parameter update result output by the first parameter evolution layer sequence is written into the corresponding first graph convolutional layer sequence, and the layer parameter update result output by the second parameter evolution layer sequence is written into the corresponding second graph convolutional layer sequence, thereby generating a parallel connection structure of stable knowledge evolution unit and state knowledge evolution unit. A parameter interaction path is set between the first parameter evolution layer sequence and the second parameter evolution layer sequence. The layer parameter update results output by each layer in the first parameter evolution layer sequence are written into the input end of the corresponding layer in the second parameter evolution layer sequence, and the layer parameter update results output by each layer in the second parameter evolution layer sequence are written into the input end of the corresponding layer in the first parameter evolution layer sequence, thereby generating a bidirectional parameter connection structure between the stable knowledge evolution unit and the state knowledge evolution unit. Based on the stable knowledge evolution unit and the state knowledge evolution unit, the basic structure of the original EvolveGCN model is extended at the output end. A support conflict correction unit is added after the stable knowledge evolution unit and the state knowledge evolution unit, and the output ends of the first graph convolutional layer sequence and the second graph convolutional layer sequence are connected to the input end of the support conflict correction unit. Based on the event unfolding encoding unit, stable knowledge evolution unit, state knowledge evolution unit, and supporting conflict correction unit, as well as the connection relationships between each unit, a dual-evolution EvolveGCN model is generated. During the training of the dual-evolutionary EvolveGCN model, historical emergency and critical care clinical guidelines, expert consensus data, historical reference data, and historical emergency and critical care treatment data were collected. Field splitting, cleaning, time alignment, terminology standardization, and structured organization were performed on these data to generate a historical emergency and critical care medical knowledge source data sequence. Based on this sequence, a historical emergency and critical care treatment event unfolding diagram and a historical time slice diagram sequence were constructed. For each historical emergency and critical care medical knowledge source data sequence, the treatment object node, clinical event node, evidence status node, chronological relationship, evidence support relationship, and treatment trigger relationship were extracted from each time slice according to the time index field, forming the model training input samples. Then, based on... The system constructs labels for each relational connection by considering the continuous occurrence of the same historical emergency medical knowledge source data sequence in subsequent time slices, the medical description results corresponding to the description content of the treatment object, the medical description results corresponding to the description content of the clinical event, the evidence description results corresponding to the description content of the evidence status, the confirmation results corresponding to the final diagnosis data, the execution results corresponding to the treatment data, and the consistency of the source label field. Among them, relational connections that exist in multiple consecutive time slices and are consistent with the final diagnosis data and treatment data are marked as stable knowledge labels; relational connections that are added, disappear, or change in intensity between adjacent time slices are marked as state knowledge labels; and relational connections that are continuously supported in subsequent time slices and do not have conflict records are marked as correction relational labels. This forms training labels that correspond one-to-one with the model training input samples. The training input samples are fed into the dual-evolution EvolveGCN model in time slice order. First, the event expansion encoding unit encodes the node connection structure and relation connection structure in each time slice to obtain the event expansion representation corresponding to each time slice. Then, the event expansion representation is simultaneously input into the stable knowledge evolution unit and the state knowledge evolution unit. The stable knowledge evolution unit outputs the stable relation prediction results corresponding to each time slice, and the state knowledge evolution unit outputs the state relation prediction results corresponding to each time slice. Afterward, the stable relation prediction results and the state relation prediction results are jointly input into the support conflict correction unit, which outputs the corrected relation prediction results corresponding to each relation connection. The stable relation prediction results, state relation prediction results, and corrected relation prediction results are all arranged according to time slice, node pair, and relation connection category to form a prediction result set corresponding to the training labels. The first loss is constructed based on the stable relation prediction results and stable knowledge tags. The specific calculation method is as follows: For each relation connection in each time slice, the predicted value of that relation connection in the stable relation prediction results is read, and the tag value of that relation connection in the stable knowledge tags is read. When the tag value is 1, the negative logarithm of the predicted value of that relation connection is calculated; when the tag value is 0, the negative logarithm of the complement of the predicted value of that relation connection with respect to 1 is calculated. The summation is applied to all relation connections in all time slices and divided by the total number of relation connections to obtain the first loss. The second loss is constructed based on the state relation prediction results and state knowledge tags, and the calculation process is the same as the first loss. The first loss is the same, the difference being that the input prediction result is the state relation prediction result, and the input label is the state knowledge label; the third loss is constructed based on the corrected relation prediction result and the corrected relation label. The specific calculation method is as follows: for each relation connection, read the predicted value of the relation connection in the corrected relation prediction result, and read the label value of the relation connection in the corrected relation label. When the corrected relation label is 1, calculate the negative logarithm of the predicted value of the relation connection. When the corrected relation label is 0, calculate the negative logarithm of the complement of the predicted value of the relation connection with respect to 1. Summing all relation connections and dividing by the total number of relation connections, the third loss is obtained. Based on the changes in event unfolded representations between adjacent time slices, a temporal continuous loss is constructed. Specifically, for the event unfolded representations corresponding to two adjacent time slices in the same historical emergency medical knowledge source data sequence, the difference is calculated dimension by dimension. Then, the difference is squared and summed to obtain the representation change value of the adjacent time slice pair. The representation change values ​​of all adjacent time slice pairs are averaged to obtain the temporal continuous loss. On this basis, the first loss, second loss, third loss and temporal continuous loss are weighted and summed to generate the total loss function. The weight coefficients of the first loss, second loss, third loss and temporal continuous loss are set according to the loss reduction rate on the validation sample set. Subsequently, the Adam optimizer was used to iteratively update the parameters in the dual-evolution EvolveGCN model. The initial learning rate was set to 0.001 and the batch size was set to 32. After each training epoch, the validation loss corresponding to the total loss function was calculated on the validation sample set. When the decrease in validation loss was less than 0.0001 for 10 consecutive training epochs, or when the number of training epochs reached 100, the dual-evolution EvolveGCN model was deemed to have reached the convergence condition. The model parameters corresponding to the minimum validation loss were saved, and the trained dual-evolution EvolveGCN model was obtained.

[0024] In this embodiment, generating stable knowledge subgraph sequences and state knowledge subgraph sequences includes: Based on the time slice graph sequence, the diagnosis and treatment object nodes, clinical event nodes, evidence status nodes, and corresponding relationship connections in each time slice are extracted according to the time slice order corresponding to the time index field. The diagnosis and treatment object identifier in the diagnosis and treatment object node is combined with the time index field to form the diagnosis and treatment object node input vector. The clinical event marker, time index field, and source marker field in the clinical event node are combined to form the clinical event node input vector. The evidence status marker, time index field, and source marker field in the evidence status node are combined to form the evidence status node input vector. According to the chronological order of time slices, the input vectors of the diagnosis and treatment object nodes, the input vectors of the clinical event nodes, the input vectors of the evidence status nodes, and the relational connections corresponding to each time slice are written into the event expansion encoding unit. Specifically, within each time slice, the input vectors of the diagnosis and treatment object nodes, the input vectors of the clinical event nodes, and the input vectors of the evidence status nodes are written first, and then the chronological order of occurrence, the evidence support relationship, and the diagnosis and treatment trigger relationship are written. Based on the event-expanded encoding unit, encoding processing is performed on the input vectors of the treatment object node, clinical event node, and evidence status node within each time slice, along with their corresponding relational connections. This yields node representations and relational representations corresponding to each time slice. Node representations show the encoding results of the corresponding treatment object node, clinical event node, and evidence status node within the current time slice, while relational representations show the encoding results of the sequence of occurrence, evidence support, and treatment triggering relationships within the current time slice. Specifically, the encoding processing involves mapping the input vectors of the treatment object node, clinical event node, and evidence status node to a unified... The initial node representation of each dimension is generated. A relational connection index is constructed within the current time slice based on the order of occurrence, evidence support, and treatment triggering. A relational category label and connection direction label are written for each relational connection. Based on the relational connection index, the initial node representations of adjacent nodes directly connected to each node are extracted. Aggregation processing is performed according to the relational category label and connection direction label, and the aggregation result is combined with the initial node representation of the current node to generate the node representation corresponding to the current time slice. Finally, a combined encoding process is performed based on the node representations of the nodes at both ends of each relational connection, the relational category label, and the connection direction label to generate the relational representation corresponding to the current time slice. Based on the node representations and relation representations in each time slice, the aggregation process is performed according to the time slice order. The node representations and relation representations in the same time slice are combined to generate the time slice representation corresponding to the current time slice. The time slice representations corresponding to each time slice are arranged in chronological order to generate the event expansion representation. Based on event expansion representation, relational separation processing is performed on relational connections in each time slice. Specifically, for the same relational connection, the persistence of the relational connection in consecutive time slices, the consistency of the source marker field, and the change of the evidence status marker are statistically analyzed. When the same relational connection persists in consecutive time slices, the source marker field remains consistent, and the evidence status marker does not change according to the corresponding relational category, the relational connection is determined to be a relational connection reflecting a long-term medical relationship. When the same relational connection is added, interrupted, has a change in the connection object, or has a change in the evidence status marker between adjacent time slices, the relational connection is determined to be a relational connection reflecting a change in the emergency treatment process. Based on the relational connections reflecting long-term medical relationships and the corresponding treatment object nodes, clinical event nodes, and evidence state nodes in the time slices, a stable knowledge subgraph corresponding to each time slice is constructed. The stable knowledge subgraphs are then organized in the order of the time slices to generate a stable knowledge subgraph sequence. Based on the relational connections reflecting changes in the emergency treatment process and the corresponding treatment object nodes, clinical event nodes, and evidence state nodes in the time slices, a state knowledge subgraph corresponding to each time slice is constructed. The state knowledge subgraphs are then organized in the order of the time slices to generate a state knowledge subgraph sequence. The stable knowledge subgraph sequence stores the long-term medical relationship connection structure within each time slice in the order of the time slices, while the state knowledge subgraph sequence stores the relationship connection structure of changes in the emergency treatment process within each time slice in the order of the time slices.

[0025] In this embodiment, generating a union representation includes: Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the stable knowledge subgraph corresponding to each time slice is input into the stable knowledge evolution unit in the order of time slices, and the state knowledge subgraph corresponding to each time slice is input into the state knowledge evolution unit. The node connection structure and relation connection structure in the current time slice are extracted in the stable knowledge evolution unit as the current input of the first parameter evolution process, and the node connection structure and relation connection structure in the current time slice are extracted in the state knowledge evolution unit as the current input of the second parameter evolution process. In the stable knowledge evolution unit, the first parameter evolution processing is performed on the current input according to the time slice order. Specifically, the evolution result of the relation connection structure corresponding to the previous time slice is used as the initial parameter basis for the current time slice. Combining the node connection structure and relation connection structure in the current time slice, the relation connection parameters and node propagation parameters in the current time slice are updated to generate the stable relation representation corresponding to the current time slice. Based on the relation connections that maintain consistency in relation category, connection object, and source label field in consecutive time slices, the medical relation boundary corresponding to the current time slice is extracted. The medical relation boundary is the stable relation constraint result extracted from the relation connections that maintain consistency in relation category, connection object, and source label field in consecutive time slices based on the stable relation representation. The medical relation boundary is used to record the range of relation connections that can be retained as stable medical relations in the current time slice. In the state knowledge evolution unit, the second parameter evolution process is performed on the current input according to the time slice order. Specifically, the evolution result of the relation connection structure corresponding to the previous time slice is used as the initial parameter basis for the current time slice. Combined with the node connection structure and relation connection structure in the current time slice, the relation connection parameters and node propagation parameters in the current time slice are updated to generate the preliminary time-varying relation representation corresponding to the current time slice. Based on the newly added relation connections, interrupted relation connections, relation connections whose connection objects have changed, and relation connections whose evidence state markers have changed between adjacent time slices, the time-varying relation representation corresponding to the current time slice is generated. Within each time slice, the time-varying relation representation corresponding to the current time slice is introduced into the relation connection parameter update stage of the first parameter evolution process. Specifically, the result reflecting the intensity of relation connection change in the time-varying relation representation is written into the parameter update process of each relation connection in the current time slice. For relation connections with change intensity higher than a preset threshold, the stable propagation weight in the current time slice is reduced, and for relation connections with change intensity lower than a preset threshold, the stable propagation weight in the current time slice is retained, thereby generating a stable relation representation after cross-introduction of the time-varying relation representation. Within each time slice, the medical relationship boundary corresponding to the current time slice is introduced into the relationship connection parameter update stage of the second parameter evolution process. Specifically, based on the medical relationship boundary, it is determined whether each relationship connection within the current time slice is within the range of medical relationships that remain stable in consecutive time slices. The state change propagation weight is reduced for relationship connections within the medical relationship boundary, and the state change propagation weight is retained for relationship connections outside the medical relationship boundary, thus generating a time-varying relationship representation after the cross-introduction of the medical relationship boundary. In the stable knowledge evolution unit and the state knowledge evolution unit, according to the correspondence of the same time slice, the same relation category, and the same connection object, the stable relation representation after the cross-introduction of time-varying relation representation and the time-varying relation representation after the cross-introduction of medical relation boundary are aligned and combined to generate the joint relation representation corresponding to the current time slice; The first parameter evolution process, the second parameter evolution process, the time-varying relation representation introduction process, the medical relation boundary introduction process, and the alignment combination process are repeatedly executed in time slice order to generate joint relation representations corresponding to the stable knowledge subgraph sequence and the state knowledge subgraph sequence.

[0026] In this embodiment, generating the set of correction relationships includes: Based on the union relation representation, the relation connections, relation categories, connection objects, time index fields, source tag fields, and evidence status tags within each time slice are extracted in time slice order, and the union relation representation corresponding to each relation connection is determined as the relation representation of the corresponding relation connection; Within the same time slice, support determination is performed on each pair of relational connections. When two relational connections have the same connection object, the same relational category, the time index field corresponds to the same time slice, the source mark field corresponds to different sources, and the evidence status marks corroborate each other, the two relational connections are constructed as a connection pair in the support relational graph, and the corresponding relational representation is configured on the connection pair to generate the support relational graph. Within the same time slice, conflict determination is performed on each pair of relational connections. When the connection objects of two relational connections are consistent and meet any of the preset specific conditions, the two relational connections are constructed as a connection pair in the conflict relation graph, and the corresponding relation representation is configured on the connection pair to generate the conflict relation graph. The preset specific conditions include: the relation categories of the two relational connections are contradictory, the time index fields of the two relational connections exceed the allowed range of the order of diagnosis and treatment, and the record sources corresponding to the source mark fields of the two relational connections are inconsistent and the evidence status marks are incompatible. The support relationship graph and the conflict relationship graph are input into the support conflict correction unit. Support propagation processing is performed on the relationship representations in the support relationship graph according to the connection direction within the graph, and conflict propagation processing is performed on the relationship representations in the conflict relationship graph according to the connection direction within the graph. Specifically, the support propagation processing is based on the relationship representations of the other relationships connected to the current relationship in the support relationship graph, and performs cumulative updates on the relationship representations of the current relationship to generate the support propagation result. Similarly, the conflict propagation processing is based on the relationship representations of the other relationships connected to the current relationship in the conflict relationship graph, and performs cumulative updates on the relationship representations of the current relationship to generate the conflict propagation result. Based on the support propagation results and conflict propagation results, a dual-graph merging update is performed on each relation connection within the same time slice. The update result corresponding to the current relation connection in the support propagation results is used as the retained update part, and the update result corresponding to the current relation connection in the conflict propagation results is used as the suppressed update part. The relation representation of the current relation connection is updated by difference, and the initial corrected relation representation corresponding to the current time slice is generated. Based on the consistency of connected objects and relation categories between adjacent time slices, a reverse correction process is performed on the initial corrected relation representation in the next time slice. Specifically, the initial corrected relation representation of the corresponding relation connection in the previous time slice is written into the correction input of the corresponding relation connection in the next time slice. The relation connection corresponding result suppressed by the conflict propagation process in the previous time slice is used as the suppression basis for relation connections with the same connected object and relation category in the next time slice. The relation connection corresponding result retained by the support propagation process in the previous time slice is used as the continuation basis for relation connections with the same connected object and relation category in the next time slice. The initial corrected relation representation in the next time slice is updated again to generate a cross-time slice corrected relation representation. According to the time slice order, the cross-time slice correction relationship representations in each time slice are summarized, and the relationship connections retained after double graph propagation processing, the relationship connections that continue to be retained after reverse correction processing, the relationship connections that are suppressed after double graph propagation processing, and the relationship connections that are removed after reverse correction processing are marked respectively to generate a correction relationship set.

[0027] In this embodiment, generating an emergency medical knowledge base includes: Based on the set of corrected relations, the time index field, relation category, connection object, source tag field, correction result and union relation representation corresponding to each relation connection are extracted according to the time slice order of the relation connection. The relation connections corresponding to the same connection object are arranged in order according to the time index field to generate the database entry judgment sequence. For each relation connection in the data entry decision sequence, a time consistency check is performed. Specifically, the time index field of the current relation connection is compared with the time index field of the relation connection of the same object in the previous time slice. If the time index field of the current relation connection is later than the time index field of the relation connection in the previous time slice, and no time rollback conflict occurs, the current relation connection is determined to satisfy time consistency. If the time index field of the current relation connection is earlier than the time index field of the relation connection in the previous time slice, or if the current relation connection crosses the boundary of the established diagnosis and treatment order in the emergency medical knowledge source data sequence, the current relation connection is determined to not satisfy time consistency. For each relation connection in the inbound judgment sequence, a relation consistency judgment is performed. Specifically, the relation category expression, the connection object expression in the joint relation representation, and the relation correction status in the correction result are extracted for the current relation connection. The current relation connection is compared with the historical corrected relation connections corresponding to the same connection object one by one. When the relation category is consistent, the connection object is consistent, and the relation correction status does not show a negative result supporting the conflict correction, the current relation connection is determined to satisfy relation consistency. When the relation category conflicts, the connection object is mismatched, or the relation correction status corresponds to a conflict retention result, the current relation connection is determined to not satisfy relation consistency. For each relational connection in the entry judgment sequence, a source consistency judgment is performed. Specifically, the source marker field corresponding to the current relational connection is extracted, and the source marker field is matched with the source record in the evidence status node corresponding to the current relational connection. The source marker fields of the current relational connection are also cross-compared with those of other relational connections corresponding to the same connection object. When the source marker field points to the same source record link and the source marker field is consistent with the source record in the evidence status node, the current relational connection is determined to satisfy source consistency. When the source marker field has any of the following conditions: source break, source conflict, or source missing, the current relational connection is determined to not satisfy source consistency. For current relational connections that meet the time consistency, relational consistency, and source consistency requirements, they are determined to meet the entry conditions. A formal relational record area is constructed, and the current relational connection, the corresponding connection object, relational category, time index field, source tag field, and correction result are written into the formal relational record area to generate an entry relational connection. For current relational connections that do not meet the requirements of time consistency, relational consistency, and source consistency, perform verification marking processing, construct a relational record area to be verified, and write the current relational connection along with the corresponding connection object, relational category, time index field, source mark field, correction result and failed consistency judgment item into the relational record area to be verified, thus generating a relational connection to be verified; Based on the inbound relationship connections and corresponding connection objects, relationship categories, time index fields, source marker fields, and correction results in the formal relationship record area, and based on the unverified relationship connections and corresponding connection objects, relationship categories, time index fields, source marker fields, correction results, and failed consistency judgment items in the unverified relationship record area, an emergency medical knowledge base is generated.

[0028] A system for constructing an emergency medical knowledge base based on medical knowledge graphs includes: The emergency medical knowledge source data processing module is used to collect emergency and critical care clinical guideline data, expert consensus data, reference data, and emergency treatment data. The emergency treatment data includes chief complaint data, vital sign data, laboratory data, imaging data, diagnostic data, treatment data, and time-stamped data. Based on unified field rules, unified time index rules, and unified medical terminology mapping rules, the module performs cleaning, alignment, and standardization processing on the emergency and critical care clinical guideline data, expert consensus data, reference data, and emergency treatment data. It then combines patient identification, clinical event markers, and evidence status markers to structure and organize the data, generating an emergency medical knowledge source data sequence. The time slice diagram construction module is used to establish treatment object nodes, clinical event nodes, and evidence status nodes based on the emergency medical knowledge source data sequence. It connects the chronological order of occurrence, evidence support relationship, and treatment trigger relationship between the treatment object nodes, clinical event nodes, and evidence status nodes to construct an emergency medical event unfolding diagram. Based on the time distribution of each node and relationship connection in the emergency medical event unfolding diagram, it generates a time slice diagram sequence. The dual-evolution EvolveGCN model construction module is used to call the basic structure of the original EvolveGCN model and rewrite the basic structure of the original EvolveGCN model based on the emergency medical event unfolding diagram and time slice diagram sequence to generate the dual-evolution EvolveGCN model. The dual-evolution EvolveGCN model includes an event unfolding encoding unit, a stable knowledge evolution unit, a state knowledge evolution unit, and a support conflict correction unit. The output of the event unfolding encoding unit is connected to the stable knowledge evolution unit and the state knowledge evolution unit respectively. A parameter interaction path is set between the stable knowledge evolution unit and the state knowledge evolution unit. The outputs of the stable knowledge evolution unit and the state knowledge evolution unit are connected to the support conflict correction unit. The event expansion separation module is used to input the diagnosis and treatment object nodes, clinical event nodes, and evidence status nodes in the emergency diagnosis and treatment event expansion graph into the event expansion encoding unit based on the time slice graph sequence. It performs encoding processing on the node connection structure and relation connection structure in each time slice according to the time slice order to generate the event expansion representation. Based on the event expansion representation, it separates the relation connections reflecting long-term medical relationships from the relation connections reflecting changes in the emergency process to generate a stable knowledge subgraph sequence and a state knowledge subgraph sequence. The joint relation evolution module is used to perform first parameter evolution processing and second parameter evolution processing in the stable knowledge evolution unit and the state knowledge evolution unit, respectively, based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence. In the first parameter evolution processing, the time-varying relation representation obtained by the second parameter evolution processing is introduced, and in the second parameter evolution processing, the medical relation boundary obtained by the first parameter evolution processing is introduced. The relation connections corresponding to the stable knowledge subgraph sequence and the state knowledge subgraph sequence are updated in a linked manner to generate the joint relation representation. The support conflict correction module is used to construct a support relationship graph based on the joint relationship representation, and to construct a conflict relationship graph based on the contradictory, time-invalid, and source-inconsistent relationship connections. The support relationship graph and the conflict relationship graph are input into the support conflict correction unit, and the dual-graph propagation processing is performed within the same time slice, and the reverse correction processing is performed between adjacent time slices to generate a set of corrected relationships. The inbound judgment module is used to perform inbound condition judgment on the relation connections in the corrected relation set according to time consistency, relation consistency, and source consistency. For relation connections that meet the inbound conditions, a formal relation record area is constructed, and the corresponding connection object, relation category, time index field, source tag field, and correction result are written into the formal relation record area to generate the inbound relation connection. For relation connections that do not meet the inbound conditions, a relation record area to be verified is constructed, and the corresponding connection object, relation category, time index field, source tag field, correction result, and failed consistency judgment items are written into the relation record area to be verified to generate the relation connection to be verified. The emergency medical knowledge base generation module is used to generate an emergency medical knowledge base based on the inbound relationship connections and corresponding connection objects, relationship categories, time index fields, source marker fields and correction results in the formal relationship record area, and based on the unverified relationship connections and corresponding connection objects, relationship categories, time index fields, source marker fields, correction results and failed consistency judgment items in the unverified relationship record area.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the scenario of organizing and updating emergency medical knowledge in a general hospital. In this scenario, the emergency and first-aid related information received daily comes not only from electronic medical records, but also from emergency and first-aid clinical guidelines, expert consensus data, reference data, and emergency medical data. Information from different sources differs significantly in terms of expression, recording granularity, time stamping methods, and terminology. Emergency medical data can reflect the actual treatment process, but its coverage is limited by a single patient visit; clinical guidelines, expert consensus data, and reference data can provide relatively stable boundaries of medical knowledge, but they suffer from inconsistencies in expression, discontinuous timelines, and inconsistent evidence status compared to the actual emergency process. Current practices typically employ three methods to complete knowledge entry into the database: manual sorting, keyword retrieval plus manual input, and static rule extraction. These methods are prone to problems such as duplicate recording of the same patient, inconsistent descriptions of the same clinical event in different sources, mixing of evidence-supporting relationships and treatment-triggered relationships, and failure to effectively remove expired relationships. As a result, the knowledge base is updated slowly, the reliability of relationships is unstable, and subsequent search and retrieval results are not reliable enough.

[0030] In this embodiment, the collected input data includes four types of sources: 126 emergency and critical care clinical guidelines, 84 expert consensus documents, 312 references, and 1860 continuous emergency treatment records. The emergency treatment data includes chief complaint data, vital signs data, laboratory data, imaging data, diagnostic data, treatment data, and time-stamped data. After processing the data using unified field rules, corresponding descriptions of the treatment object, clinical events, evidence status, time index fields, and source marker fields are generated. Considering that guidelines, consensus documents, and literature primarily use itemized, paragraph-based, and conclusion-based knowledge expression, while emergency treatment data primarily uses case-process-based knowledge expression, this embodiment employs unified time index rules and unified medical terminology mapping rules to organize the core expressions from different sources into a unified structure, generating an emergency medical knowledge source data sequence. After cleaning, the total number of original record entries was reduced from 23,840 to 22,176, eliminating duplicate field records, obviously abnormal field records, and records with missing key fields. After further standardizing the terminology, synonyms were merged into unified medical terminology items, resulting in 4,187 terminology mapping entries.

[0031] Based on the emergency medical knowledge source data sequence, an emergency treatment event unfolding graph was constructed, and a time slice graph sequence was generated according to the time index field. For each continuous emergency treatment process, the treatment object identifier was mapped to the treatment object node, changes in chief complaint, changes in test results, diagnosis confirmation, and treatment execution were mapped to clinical event nodes, and evidence supporting the current event was mapped to evidence status nodes. Subsequently, chronological relationships, evidence support relationships, and treatment trigger relationships were established. After graph construction, a total of 1860 treatment object nodes, 14892 clinical event nodes, 17346 evidence status nodes, 26540 chronological relationships, 21408 evidence support relationships, and 14892 treatment trigger relationships were formed. After segmentation according to the time index field, 9340 time slice graph sequences were generated. In this way, the knowledge content originally scattered in different texts and different case records was transformed into a graph structure with continuous evolution characteristics, providing a unified input basis for subsequent relationship evolution and conflict correction.

[0032] In the model processing stage, a dual-evolutionary EvolveGCN model is constructed, and the unfolded graph of emergency medical events and the time slice graph sequence are input into the event unfolding encoding unit. The model separates long-term stable medical relationships from the changing relationships in the emergency process, generating stable knowledge subgraph sequences and state knowledge subgraph sequences. The stable knowledge subgraph sequences mainly carry relatively stable relationship boundaries from guidelines, consensus, and references, while also absorbing recurring relationship connections in emergency medical data that do not change significantly with individual processes; the state knowledge subgraph sequences mainly carry relationship connections corresponding to changes in condition, changes in testing, adjustments in treatment, and changes in evidence status. Subsequently, the first parameter evolution processing and the second parameter evolution processing are performed in the stable knowledge evolution unit and the state knowledge evolution unit, respectively, and are updated in conjunction with each other during the two evolution processes to generate a joint relationship representation. Based on the joint relationship representation, mutually validated relationship connections are constructed into a supporting relationship graph, and mutually contradictory, time-invalid, and inconsistent source relationship connections are constructed into a conflict relationship graph. These are then input into the supporting conflict correction unit to perform dual-graph propagation processing and reverse correction processing to obtain a corrected relationship set.

[0033] During the data entry phase, this embodiment determines the relationship connections in the correction relationship set based on time consistency, relational consistency, and source consistency. Time consistency primarily determines whether the time index field corresponding to the relationship connection is consistent with the order of the diagnosis and treatment process, preventing erroneous relationships where a later event supports a earlier event from entering the knowledge base. Relational consistency primarily determines whether the relationship category and the connected object conform to the established relationship boundaries and the current evolution result, preventing mixed writing of supporting relationships, triggering relationships, and sequential relationships. Source consistency primarily determines whether there is a significant contradiction between the source marker field and the relationship content, preventing low-confidence sources from directly overwriting high-confidence source relationships. Relationship connections that meet the data entry conditions enter the formal relationship record area, while those that do not enter the unverified relationship record area. The final generated emergency medical knowledge base contains 19,864 formal relationship records and 2,317 unverified relationship records. The unverified relationship records retain the source marker field, correction results, and failed consistency judgment items for subsequent manual confirmation and incremental updates. Compared with traditional manual organization methods, the method of this invention does not pursue an excessively high automatic entry rate. Instead, it retains some disputed relationship connections in the relationship record area to be verified, ensuring that the overall quality of the knowledge base is more robust and meets the requirements of knowledge credibility in emergency medical scenarios.

[0034] To verify the practical effect of this invention, this embodiment sets up three methods for comparison. Method 1 is the traditional manual data entry method, in which medical information organizers manually extract, judge, and enter data based on emergency medical data, clinical guidelines, and literature. Method 2 is a rule-template-based knowledge extraction method, which extracts and enters data through preset field rules, relational word rules, and a manually maintained terminology mapping table. Method 3 is the method of this invention, namely, the emergency medical knowledge base construction method based on medical knowledge graphs. The comparative experiments used the same batch of datasets, and the evaluation indicators selected were six indicators: knowledge relationship accuracy, conflict relationship identification rate, time-invalidated relationship removal rate, formal data entry pass rate, unverified relationship retention rate, and single batch construction time. Relevant data are shown in Table 1.

[0035] Table 1 Comparison of the Construction Effects of Emergency Medical Knowledge Base

[0036] As shown in Table 1, the method of this invention achieves a knowledge relationship accuracy of 90.7%, which is 6.1 percentage points higher than traditional manual organization methods and 10.9 percentage points higher than rule-template-based knowledge extraction methods. This result indicates that the method of this invention does not rely solely on static field matching to establish relationship connections. Instead, it generates emergency medical event unfolding diagrams and time-slice diagram sequences based on emergency medical knowledge source data sequences. Within a temporal evolution framework, it organizes treatment object nodes, clinical event nodes, and evidence state nodes, thus better distinguishing between truly effective relationship connections and superficial co-occurrence relationship connections. For continuous process information such as frequent test result updates, diagnostic corrections, and treatment adjustments in emergency scenarios, this graphical representation can reduce erroneous associations.

[0037] In terms of conflict relationship identification rate and time-invalidated relationship removal rate, the method of this invention achieves 88.2% and 86.4% respectively, demonstrating a significant advantage. This result is directly related to the joint correction mechanism of the support relationship graph and the conflict relationship graph. Traditional manual sorting methods are affected by the experience and workload of the sorting personnel, and are prone to overlooking implicit contradictions between sources; although rule template-based knowledge extraction methods are fast, their ability to identify cross-time-slice relationship reversals, evidence invalidation, and source conflicts is limited. The method of this invention performs dual-graph propagation processing and reverse correction processing on the basis of joint relationship representation, which can incorporate relationship support and relationship conflicts within the same time slice into a unified processing flow, while continuously correcting relationship connections in adjacent time slices. Therefore, the identification of time-invalidated relationships and source inconsistency relationships is more stable.

[0038] Regarding the formal database entry pass rate, the method of this invention achieves 89.1%, higher than the other two methods. However, simply pursuing a high pass rate can easily lead to the direct writing of relationships with unclear boundaries into the formal relationship record area, thus increasing the subsequent error correction cost of the knowledge base. The method of this invention retains a 10.4% retention rate for relationships to be verified, which is slightly higher than the knowledge extraction method based on rule templates. This indicates that the method of this invention adopts a more conservative approach to handling controversial relationship connections. This result does not represent a decrease in capability, but rather that the system can identify relationship connections that are not suitable for direct database entry and retain them in the relationship record area to be verified with evidence, preventing unreliable knowledge from directly entering the formal relationship record area. For medical scenarios, this result is more in line with practical application requirements.

[0039] In terms of construction time, the method of this invention takes 5.1 hours per batch, which is lower than the 19.6 hours of traditional manual processing methods and also lower than the 7.8 hours of rule template-based knowledge extraction methods. This is because the method of this invention does not perform extensive manual verification in the later stages. Instead, it uses layered processing during the relation generation stage, employing emergency medical event unfolding graphs, stable knowledge subgraph sequences, state knowledge subgraph sequences, and corrected relation sets to preemptively address many issues that could have surfaced after data entry. As a result, the relation connections in the formal relation record area are of higher quality, significantly reducing subsequent rework and compressing the overall processing cycle.

[0040] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing an emergency medical knowledge base based on medical knowledge graphs, characterized in that, Includes the following steps: Collect emergency and critical care clinical guidelines data, expert consensus data, reference data, and emergency treatment data; perform cleaning, alignment, and standardization processing to generate an emergency medical knowledge source data sequence. Based on the emergency medical knowledge source data sequence, an emergency diagnosis and treatment event unfolding graph is constructed, and a time slice graph sequence is generated based on the time distribution of each node and each relationship connection. A dual-evolution EvolveGCN model is constructed, including an event unfolding encoding unit, a stable knowledge evolution unit, a state knowledge evolution unit, and a support conflict correction unit. Based on the time slice graph sequence, the emergency medical event unfolding graph is input into the event unfolding encoding unit. The node connection structure and relation connection structure in each time slice are encoded to generate the event unfolding representation. The relation connections reflecting long-term medical relationships and the relation connections reflecting changes in the emergency process are separated to generate a stable knowledge subgraph sequence and a state knowledge subgraph sequence. Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the first parameter evolution process and the second parameter evolution process are performed in the stable knowledge evolution unit and the state knowledge evolution unit, respectively, to generate a joint relation representation; Based on the joint relation representation, a support relation graph and a conflict relation graph are constructed, and the support conflict correction unit is input to perform dual graph propagation processing and reverse correction processing to generate a set of correction relations; Based on the set of correction relationships, the entry conditions are judged to generate an emergency medical knowledge base; The construction of the dual-evolution EvolveGCN model includes: The basic structure of the original EvolveGCN model is invoked, including a sequence of graph convolutional layers that process graph sequences in time order and a sequence of parameter evolution layers connected to the graph convolutional layer sequence. Based on the emergency medical event unfolding diagram and time slice sequence, the basic structure of the original EvolveGCN model is rewritten at the input end, and an event unfolding encoding unit is added. Based on the graph convolutional layer sequence and parameter evolution layer sequence, a two-branch rewriting is performed to split the single-path time evolution structure into a stable knowledge evolution unit and a state knowledge evolution unit. Based on the stable knowledge evolution unit and the state knowledge evolution unit, the basic structure of the original EvolveGCN model is extended at the output end, and a support conflict correction unit is added. Based on the event unfolding encoding unit, stable knowledge evolution unit, state knowledge evolution unit, and supporting conflict correction unit, as well as the connection relationships between each unit, a dual-evolution EvolveGCN model is generated. The generated union relation representation includes: Based on the stable knowledge subgraph sequence and the state knowledge subgraph sequence, the stable knowledge subgraph corresponding to each time slice is input into the stable knowledge evolution unit, and the state knowledge subgraph corresponding to each time slice is input into the state knowledge evolution unit. In the stable knowledge evolution unit, the first parameter evolution processing is performed on the current input according to the time slice order to generate a stable relation representation corresponding to the current time slice and extract the medical relation boundary; In the state knowledge evolution unit, the second parameter evolution process is performed on the current input in the order of time slices to generate a time-varying relation representation; Within each time slice, the time-varying relation representation corresponding to the current time slice is introduced into the relation connection parameter update stage of the first parameter evolution process to generate a stable relation representation after cross-introduction of the time-varying relation representation; Within each time slice, the medical relationship boundary corresponding to the current time slice is introduced into the relationship connection parameter update stage of the second parameter evolution process to generate a time-varying relationship representation after the cross-introduction of the medical relationship boundary; In the stable knowledge evolution unit and the state knowledge evolution unit, the stable relation representation after the cross-introduction of time-varying relation representation and the time-varying relation representation after the cross-introduction of medical relation boundary are aligned and combined to generate a joint relation representation.

2. The method for constructing an emergency medical knowledge base based on medical knowledge graphs according to claim 1, characterized in that, The generated emergency medical knowledge source data sequence includes: Data from emergency and critical care clinical guidelines, expert consensus, references, and emergency treatment data were collected. The emergency treatment data included chief complaint data, vital signs data, laboratory data, imaging data, diagnostic data, treatment data, and time-stamped data. Following unified field rules, field splitting was performed on emergency and critical care clinical guideline data, expert consensus data, reference data, and emergency treatment data to generate a basic record set. Based on the unified time indexing rules, time alignment processing is performed on the basic record set to generate a time-aligned record set; Based on the unified medical terminology mapping rules, terminology standardization processing is performed on the time-aligned record set to generate a standardized record set; Based on a standardized record set, according to the data ownership relationship in the same continuous emergency process, the same medical object is assigned an identifier, written into the corresponding data record, and an object-marked record set is generated. Based on the object-tagged record set, each data is mapped to a corresponding clinical event, each clinical event is assigned a clinical event tag, written to the corresponding data record, and an event-tagged record set is generated. Based on the event marker record set, merge processing is performed within the same time slice to generate evidence status markers, write the corresponding data records, and generate an evidence marker record set; Based on the evidence-tagged record set, a structured organization is performed to generate a sequence of emergency medical knowledge source data.

3. The method for constructing an emergency medical knowledge base based on medical knowledge graphs according to claim 1, characterized in that, The generated time slice sequence includes: Based on the emergency medical knowledge source data sequence, the fields of diagnosis and treatment object identifier, clinical event marker, evidence status marker and source marker are read one by one to establish diagnosis and treatment object node, clinical event node and evidence status node; Write the treatment object identifier and time index field into the treatment object node, write the clinical event marker, time index field and source marker field into the clinical event node, and write the evidence status marker, time index field and source marker field into the evidence status node to form the node attribute fields of the treatment object node, clinical event node and evidence status node; Establish the chronological order of occurrence based on the clinical event nodes corresponding to the same patient identification; Establish evidence support relationships based on the evidence status nodes and clinical event nodes corresponding to the same diagnosis and treatment object identifier; Establish a diagnosis and treatment trigger relationship based on the diagnosis and treatment object node and clinical event node corresponding to the same diagnosis and treatment object identifier; Based on the diagnosis and treatment object node, clinical event node, evidence status node, sequence of occurrence, evidence support relationship and diagnosis and treatment trigger relationship, construct an emergency diagnosis and treatment event unfolding diagram; Based on the time distribution of clinical event nodes in the emergency medical event unfolding diagram, time slices are divided according to the occurrence point of the clinical event node. Based on the temporal order between adjacent time slices, the diagnosis and treatment object nodes in the previous time slice are inherited into the next time slice. The clinical event nodes and evidence status nodes in the corresponding time slices of the next time slice are retained in the next time slice. Based on the inherited diagnosis and treatment object nodes, the clinical event nodes and evidence status nodes in the next time slice, the corresponding chronological order, evidence support, and diagnosis and treatment triggering relationships are re-established to generate a time slice sequence.

4. The method for constructing an emergency medical knowledge base based on medical knowledge graphs according to claim 1, characterized in that, The generated stable knowledge subgraph sequence and state knowledge subgraph sequence include: Based on the time slice graph sequence, the diagnosis and treatment object nodes, clinical event nodes, evidence status nodes and the corresponding relationship connections of each node are extracted from each time slice and combined to form the diagnosis and treatment object node input vector, clinical event node input vector and evidence status node input vector. According to the chronological order of the time slices, the input vectors of the diagnosis and treatment object nodes, the input vectors of the clinical event nodes, the input vectors of the evidence state nodes, and the relational connections are written into the event expansion encoding unit, and the encoding process is performed to obtain the node representation and relational representation corresponding to each time slice; Based on the node representations and relationship representations in each time slice, aggregation processing is performed to combine and generate the time slice representation corresponding to the current time slice, and then arranged in chronological order to generate the event expansion representation; Based on the event unfolding representation, relation separation processing is performed on the relational connections in each time slice to determine the relational connections that reflect long-term medical relationships and the relational connections that reflect changes in the emergency process. Based on relational connections reflecting long-term medical relationships and corresponding time slice nodes of the diagnosis and treatment object, clinical event, and evidence status, a stable knowledge subgraph sequence is constructed. Based on relational connections reflecting changes in the emergency treatment process and corresponding time slice nodes of the diagnosis and treatment object, clinical event, and evidence status, a state knowledge subgraph sequence is constructed.

5. The method for constructing an emergency medical knowledge base based on medical knowledge graphs according to claim 1, characterized in that, The set of generated correction relationships includes: Based on the union relation representation, the relation connection, relation category, connection object, time index field, source tag field and evidence status tag within each time slice are extracted, and the union relation representation corresponding to each relation connection is determined as the relation representation of the corresponding relation connection; Within the same time slice, support determination is performed on each pair of relational connections to construct connection pairs in the support relation graph, and the corresponding relation representation is configured on the connection pair to generate the support relation graph. Within the same time slice, conflict determination is performed on each pair of relational connections to construct connection pairs in the conflict relation graph, and the corresponding relation representation is configured on the connection pair to generate the conflict relation graph. Input the support relationship diagram and the conflict relationship diagram into the support conflict correction unit. Perform support propagation processing on the relationship representation in the support relationship diagram according to the connection direction within the diagram, and perform conflict propagation processing on the relationship representation in the conflict relationship diagram according to the connection direction within the diagram, generating support propagation results and conflict propagation results respectively. Based on the support propagation results and conflict propagation results, a dual-graph merging update is performed on each relation connection within the same time slice to generate an initial corrected relation representation; Based on the consistent connection objects and relationship categories between adjacent time slices, reverse correction processing is performed on the initial correction relationship representation in the later time slice to generate a cross-time slice correction relationship representation; Based on the time slice order, the cross-time slice correction relationship representations in each time slice are summarized to generate a correction relationship set.

6. The method for constructing an emergency medical knowledge base based on medical knowledge graphs according to claim 1, characterized in that, The generated emergency medical knowledge base includes: Based on the set of corrected relations, extract the time index field, relation category, connection object, source tag field, correction result and joint relation representation corresponding to each relation connection, and generate an entry judgment sequence; Perform time consistency checks on each relation connection in the inbound decision sequence; Perform a relationship consistency check on each relation connection in the inbound decision sequence; Perform source consistency checks on each relation connection in the inbound decision sequence; For current relational connections that meet the time consistency, relational consistency, and source consistency requirements, they are determined to meet the entry conditions. A formal relational record area is constructed, and the current relational connection, the corresponding connection object, relational category, time index field, source tag field, and correction result are written into the formal relational record area to generate an entry relational connection. For current relational connections that do not meet the requirements of time consistency, relational consistency, and source consistency, perform verification marking processing, construct a relational record area to be verified, and write the current relational connection along with the corresponding connection object, relational category, time index field, source mark field, correction result and failed consistency judgment item into the relational record area to be verified, thus generating a relational connection to be verified; An emergency medical knowledge base is generated based on the formal relationship record area and the relationship record area to be verified.

7. A system for constructing an emergency medical knowledge base based on a medical knowledge graph, comprising the method for constructing an emergency medical knowledge base based on a medical knowledge graph as described in any one of claims 1 to 6, characterized in that, include: The emergency medical knowledge source data processing module is used to collect emergency and first aid clinical guidelines data, expert consensus data, reference data and emergency diagnosis and treatment data, perform cleaning, alignment, standardization processing and structured organization, and generate emergency medical knowledge source data sequences. The time slice graph construction module is used to establish treatment object nodes, clinical event nodes, and evidence status nodes based on the emergency medical knowledge source data sequence, construct an emergency treatment event unfolding graph, and generate a time slice graph sequence; The dual-evolution EvolveGCN model building module is used to call the basic structure of the original EvolveGCN model, perform structural rewriting, and generate the dual-evolution EvolveGCN model. The event unfolding separation module is used to perform encoding processing in the event unfolding encoding unit based on the time slice graph sequence, generate the event unfolding representation, and separate the generated stable knowledge subgraph sequence and state knowledge subgraph sequence. The joint relation evolution module is used to generate joint relation representations by performing first parameter evolution processing and second parameter evolution processing on stable knowledge subgraph sequences and state knowledge subgraph sequences, respectively. The support conflict correction module is used to construct support relationship graphs and conflict relationship graphs based on the joint relationship representation, and to perform dual-graph propagation processing and reverse correction processing to generate a set of corrected relationships; The inbound judgment module is used to perform inbound condition judgment based on the set of correction relationships, and generate inbound relationship connections and relationship connections to be verified respectively. The emergency medical knowledge base generation module is used to generate an emergency medical knowledge base based on the relationship record content in the formal relationship record area and the relationship record area to be verified.

Images