A knowledge graph-based new drug clinical evidence association analysis system

Abstract

The present application relates to the field of clinical evidence data organization and information retrieval technology, in particular to a new drug clinical evidence correlation analysis system based on a knowledge graph. The system includes a graph processor and an evidence storage. Multiple sources of clinical evidence are converted into evidence assertion nodes bound to source locations, and clinical entity nodes and double-time version nodes are associated. Relationship edges are established according to source revision dependencies and conclusion dependencies. The system compares the versions before and after the source to form version difference events, maps the change content to the corresponding evidence assertion, determines the affected path along the relationship edge, completes local reconstruction in the candidate graph branch, switches the query index after integrity verification, and saves the historical snapshot. The present application can reduce the graph reconstruction range after the evidence changes, reduce irrelevant node traversal and index repeated writing, keep the correlation results consistent with the effective evidence, and support tracing the historical evidence state according to the clinical event time and the system effective time.

Description

Technical Field

[0001] This invention relates to the field of clinical evidence data organization and information retrieval technology, specifically a new drug clinical evidence association analysis system based on knowledge graphs. Background Technology

[0002] Existing association analyses of clinical evidence for new drugs are typically based on relational databases. Data is collected from clinical trial registration documents, research papers, drug instructions, and regulatory review materials. After text extraction, field cleaning, and terminology normalization, structured records are generated according to fields such as drug, indication, trial identifier, study population, intervention, evaluation endpoint, trial results, and adverse events. The system uses drug name, disease classification code, and trial identifier to align records across sources. It establishes associations between drugs and clinical evidence through primary and foreign key joins, field mapping rules, and search criteria. Evidence lists or statistical reports are then generated according to trial stage, data source, and result type. When data content is revised, the database typically saves the updated content by overwriting the original fields or appending new records, and re-executes the established join rules to refresh the association results.

[0003] The above-mentioned processing methods suffer from a lack of calculable dependencies between changes in evidence versions and existing evidence when dealing with revisions to experimental protocols, corrections to results announcements, errata in papers, or evidence retractions. Using the entire experiment or entire paper as the recording granularity prevents localized conclusion changes from being directly mapped to corresponding evidence fragments; overwriting updates lose the temporal context of older versions, while supplementary updates may allow mutually substitutable records to participate in the retrieval simultaneously. Because the database does not record the reference relationships between association paths and specific evidence versions, the system struggles to determine which downstream results are actually affected by version changes, and can only reconnect a larger range of data or continue to retain association content generated from older evidence, resulting in inconsistencies between the association results and the current evidence status. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph-based system for analyzing clinical evidence associations for new drugs, which can solve the problems mentioned in the background.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A knowledge graph-based clinical evidence association analysis system for new drugs includes a graph processor and an evidence storage device connected to the graph processor, wherein the evidence storage device stores a clinical evidence knowledge graph. The graph processor converts multi-source clinical evidence into evidence assertion nodes, associates each evidence assertion node with a clinical entity node and its source location, and connects version nodes that record the clinical event time and the system effective time. It also establishes dependency edges between evidence assertion nodes based on source revision dependencies and conclusion dependencies. The graph processor compares the previous and subsequent versions of the source to generate version difference events, maps the version difference events to the corresponding evidence assertion nodes, determines the affected evidence association paths along the dependency edges, performs local reconstruction of the evidence association paths, and saves the historical snapshot before the update.

[0006] Preferably, the graph processor segments multi-source clinical evidence according to document paragraph boundaries and table data boundaries, and converts the segmented evidence fragments into evidence assertion nodes composed of subject-object pairs and assertion relations that limit the semantics of subject-object pairs; The clinical entity nodes include drug entity nodes and clinical research entity nodes. The subject-object pair references the clinical entity nodes. The evidence assertion nodes are bound to source anchors consisting of source identifiers and original text location identifiers, and are associated with different sources of evidence from the same clinical trial through trial identifiers.

[0007] Preferably, the version node includes a clinical event time node and a system effective time node. The graph processor connects the clinical event time nodes according to the order of events recorded in the clinical data to form a first directed version chain, and connects the system effective time node according to the order of effectiveness and ineffectiveness of the evidence assertion node in the clinical evidence knowledge graph to form a second directed version chain. Each evidence assertion node is connected to the event position in the first directed version chain and the valid interval in the second directed version chain.

[0008] Preferably, the graph processor aligns the previous and subsequent versions of the source according to the source location and assertion semantic identifier, and writes the semantic changes of the unaligned evidence content and the aligned evidence content into the version difference event; Using the evidence assertion node corresponding to the version difference event as the seed node, a continuous directed traversal is performed along the dependency edges pointing to the downstream evidence association path until a path node that does not reference the changed version or does not have a downstream dependency edge is reached, forming an affected subgraph, and the affected subgraph is determined as the local reconstruction range.

[0009] Preferably, the graph processor forms a candidate assertion set based on the test identifier, and pairs the candidate assertions according to the standard entity references of the subject-object pair and the semantic identifiers of the assertion relationship; For evidence assertion nodes that match, their respective source anchors are preserved and a coreference edge is established between the evidence assertion nodes. For evidence assertion nodes that have semantic differences after pairing, independent assertion branches are formed according to the version nodes they are connected to.

[0010] Preferably, each system effective time node in the second directed version chain is connected to the valid start or valid end point of the corresponding evidence assertion node. The graph processor receives a retrieval request containing the clinical event query time and the system effective query time. First, it filters evidence assertion nodes that are no later than the clinical event query time according to the first directed version chain. Then, it retains evidence assertion nodes whose valid intervals cover the system effective query time according to the second directed version chain. Finally, it generates a query subgraph from the retained evidence assertion nodes.

[0011] Preferably, the dependency relationship edge associates the version identifier of the upstream evidence assertion node with the path identifier of the downstream evidence association path; The graph processor reads the changed version identifier from the version difference event, generates a dependency signature consisting of a dependency version identifier for each evidence association path, writes the path node whose dependency signature contains the changed version identifier into the propagation queue, continues to traverse the downstream dependency relationship edges according to the propagation queue, and terminates the corresponding traversal branch when the dependency signature does not contain the changed version identifier, while recording the visited path identifiers.

[0012] Preferably, the graph processor retains interface nodes at the connection between the affected subgraph and the unaffected subgraph, and writes the evidence assertion nodes and dependency edges generated by local reconstruction into a candidate graph branch that is separated from the current graph. Perform referential integrity verification on the candidate graph branch and the interface node, and after the verification passes, switch the query index from the current graph branch to the candidate graph branch, and connect the current graph branch before the switch to the historical snapshot.

[0013] Preferably, the graph processor performs object semantic consistency verification on the evidence assertion nodes connected by the coreference relationship edges. When there is a mutually exclusive assertion relationship between evidence assertion nodes with the same subject-object standard entity references, a conflict node is established connected to the corresponding evidence assertion node. The conflict nodes are respectively associated with the source anchor point and version node of the corresponding evidence assertion node, and are connected to the evidence association path referencing the evidence assertion node through the dependency edge.

[0014] Preferably, the graph processor reads the dependency edges of each reconstructed evidence association path within the candidate graph branch, verifies whether the connected evidence assertion nodes have source anchors, and verifies that the corresponding version node is within the valid range of the selected query time based on the clinical event time and the system effective time. When the verification results are consistent, a switch record containing the current query index identifier and the candidate query index identifier is generated. The active index pointer is updated according to the switch record, and the active index pointer before the switch is written to the entry node of the historical snapshot.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By converting multi-source clinical data into evidence assertion nodes bound to their source locations, and connecting these nodes to version nodes that simultaneously record clinical event times and system effective times, local changes in evidence content can be limited to the corresponding assertion and its version. After comparing versions from different sources to form version difference events, these events are mapped to the changed evidence assertion nodes. The downstream evidence association path is then traversed along the relationship edges formed by source revision dependencies and conclusion dependencies, thereby identifying the affected subgraph and reconstructing only the affected subgraph. This data processing chain transforms full relation reconnection into local graph updates starting with changed assertions, reducing redundant traversals of irrelevant graph nodes and duplicate writing of unchanged index records, and ensuring that the reconstructed association results are consistent with the currently effective evidence.

[0016] 2. A dual-timeline version chain separately stores the order of clinical events and the effective period of evidence in the system. Combined with historical snapshots, it can restore the corresponding evidence query subgraph according to the specified event time and system time, avoiding the mixed invocation of assertions from different time states. Coreference edges retain independent source anchors for evidence from different sources within the same trial. Mutually exclusive assertions are associated with their respective versions through conflict nodes, enabling the tracing of source differences and version substitution relationships along the graph path. Locally reconstructed content is first written to candidate graph branches separated from the current graph. After integrity checks of source anchors, version ranges, and dependencies, the active index is switched, and the original index is written to the historical snapshot entry point, preventing incomplete graph branches from entering the retrieval process and maintaining the structural integrity of query results during the update period. Attached Figure Description

[0017] Figure 1 Flowchart for constructing a knowledge graph of clinical evidence for new drugs; Figure 2 Flowchart for the propagation of evidence version changes and local map reconstruction; Figure 3 A distribution map of semantic change markers; Figure 4 A trend chart for retaining and excluding assertions based on dual-time evidence. Detailed Implementation

[0018] In one embodiment, a knowledge graph-based clinical evidence association analysis system for new drugs includes a graph processor and an evidence storage device connected to the graph processor. The evidence storage device stores a clinical evidence knowledge graph. The graph processor converts multi-source clinical evidence into evidence assertion nodes, associates each evidence assertion node with a clinical entity node and its source location, and connects it with version nodes that record the clinical event time and the system's effective time. It establishes dependency edges between evidence assertion nodes based on source revision dependencies and conclusion dependencies. The graph processor compares the previous and subsequent versions of the source to generate version difference events, maps the version difference events to the corresponding evidence assertion nodes, determines the affected evidence association path along the dependency edges, performs local reconstruction of the evidence association path, and saves a historical snapshot before the update.

[0019] In this embodiment, the graph processor and the evidence storage device constitute a data processing environment that integrates evidence access, semantic organization, version management, and association retrieval. The graph processor receives electronic documents or structured records corresponding to clinical trial registration data, trial protocols and their revisions, clinical result announcements, research papers, drug instructions, and regulatory review data. For each source, it establishes a source record and saves the source identifier, document type, acquisition time, document publication date, content summary value, and previous version references. The evidence storage device uses an attribute graph structure to save node attributes and relation edge attributes. At the same time, it establishes an entity index from clinical entity identifier to evidence assertion node, a tracing index from source identifier to source location, and a version index from version identifier to dependency relation edge. The graph processor completes the writing of new data, comparison of existing data, and reading of association paths through the index, avoiding directly overwriting the original document content with an untraceable single field value.

[0020] Specifically, the generation of source records is based on a single identifiable data release. The graph processor calculates the content summary value of the received document and reads the test identifier, publishing entity, and publication date recorded in the document. When the source identifiers are the same but the content summary values ​​are different, the later document is registered as a new source version and linked to the previous version. When a document lacks a stable source identifier, the publishing entity identifier, test identifier, and document type are used to form a source combination key. The version order under the combination key is then determined based on the publication date. The original document is saved as a read-only object. The paragraph position, table position, and structured fields obtained after parsing all retain back references to the original object. Any evidence assertion can be located to the original fragment that forms the assertion by the source anchor point. When a document is replaced later, the old object and its parsing results are still retained.

[0021] Furthermore, clinical entity nodes in the clinical evidence knowledge graph are distinguished using stable entity identifiers. Drug entities use the standard drug name and its standardized identifier as the primary key, while clinical research entities use the trial registration identifier or internal study identifier as the primary key. Synonyms, brand names, and historical names are used as alias attributes connected to the corresponding entities. Indications, subject population conditions, intervention protocols, evaluation endpoints, and adverse events are stored as entity or attribute nodes related to the research context. When writing new data, the graph processor first queries the stable entity identifier, then eliminates homonyms through the alias mapping table and source context. Names that cannot be directly unified are established as temporary entity nodes and saved in a state pending verification. After confirmation, temporary entities are connected to standardized entities through entity substitution edges. Existing evidence assertions are not deleted but entity references are redirected to standardized entities.

[0022] In this embodiment, the evidence assertion node represents a clinical statement that can stand independently in the original data. The graph processor writes the research object or drug entity in the statement into the assertion subject reference and the clinical result, status or relation object into the assertion object reference. The semantic direction between the subject and object is defined by the relational semantic identifier. The source position is jointly determined by the document identifier, version identifier and fragment position. For continuous text, the fragment position saves the paragraph order and character range. For table content, the table order and row and column coordinates are saved. The evidence assertion node is also connected to its clinical trial and research context, so that statements about different endpoints, different populations or different observation time points in the same document form an evidence structure that is independent of each other but can be aggregated along the trial node. When a local statement changes, only the corresponding evidence assertion enters the update scope.

[0023] In one embodiment, source revision dependency represents the inheritance, modification, or replacement relationship of a subsequent version of the same source to a previous version, and conclusion dependency represents the reference relationship of the aggregated conclusion in the evidence association path to one or more evidence assertions. When generating the evidence association path, the graph processor saves the path start point, path end point, the identifier of the relational edge passed through, and the version identifier of the referenced evidence assertion, and records the correspondence between query conditions and result sets in the path nodes. When an association result is supported by multiple sources, the evidence assertion corresponding to each source is connected to the path node through independent dependency relational edges. When reading the path, the source anchor point is traced back according to the dependency relational edges. When reconstructing the path, the evidence assertion in the effective state is reselected according to the same boundary.

[0024] In a preferred embodiment, the graph processor generates assertion sets for the current version and previous version of the source, performs semantic alignment on the corresponding assertions, and registers unaligned assertions or assertions that have undergone semantic changes after alignment in the version difference event. The version difference event stores the source of change, reference to the previous version, reference to the current version, reference to the change assertion, and change type. Then, using the change assertion as the propagation starting point, all outgoing dependency edges are retrieved, and reachable evidence-related paths are marked as paths to be reconstructed. Paths that do not reference change assertions maintain their original indexes. After the paths to be reconstructed form new nodes and edges in the candidate graph branches, the active query index is replaced. The active graph entry before replacement is saved as a historical snapshot.

[0025] Please refer to Figure 1 In a preferred embodiment, the graph processor segments multi-source clinical evidence according to document paragraph boundaries and table data boundaries, and converts the segmented evidence fragments into evidence assertion nodes composed of subject-object pairs and assertion relations that limit the semantics of subject-object pairs; clinical entity nodes include drug entity nodes and clinical research entity nodes, subject-object pairs refer to clinical entity nodes, evidence assertion nodes are bound to source anchors composed of source identifiers and original text location identifiers, and are associated with different sources of evidence from the same clinical trial through trial identifiers.

[0026] Specifically, continuous text is segmented based on the heading level, paragraph boundaries, syntactic boundaries, and citation boundaries obtained from document parsing. The heading and subsequent continuous paragraphs are assigned a common chapter identifier. Syntactic fragments containing statements of experimental subjects, intervention measures, or results retain their respective paragraph identifiers. Cross-sentence references are backfilled through references to normative entities within the same paragraph. The graph processor searches forward for the nearest normative entity with the same entity type for pronouns or abbreviations. When multiple entities with the same type exist in the same paragraph, the subject-object direction of the relation words is combined to determine the reference object. Fragments that cannot be uniquely determined retain multiple candidate references and enter the verification queue. After verification, formal evidence assertions are generated. The original character range is always stored in the source anchor to support returning the original text from the assertion node.

[0027] Preferably, the table data is parsed according to the structural relationship between table title, header hierarchy, data rows, and footnotes. The header is first expanded into the complete semantic path of each data column. The data cell, its row title, and column title are combined to form the table context. If the footnote limits the population scope, statistical caliber, or time range, the footnote identifier is linked to the limited data cell. The graph processor uses the entity corresponding to the row title as a reference in the subject-object relationship, writes the evaluation index or state relationship corresponding to the column title into the assertion relationship, and converts the cell content into another reference or assertion attribute. The evidence assertion generated by the same data cell is bound to the table ordinal position, row coordinates, column coordinates, and footnote reference. When the table rows and columns move subsequently, the source location is relocated through the semantic path of the header and the cell content.

[0028] Furthermore, the construction of subject-object pairs employs a collaborative process of canonical dictionary matching, syntactic dependency relations, and relation templates. The canonical dictionary provides candidate identifiers for drug entities and clinical research entities. Syntactic dependency relations determine the agent and patient directions of entities in statements. Relation templates map expressions representing support, negation, substitution, revision, or conflict to unified relational semantic identifiers. After performing aliasing and normalization on candidate entities, the graph processor merges different text expressions under the same canonical entity identifier into entity references without merging source anchors. For fragments containing negation ranges, the boundary of negation is preserved, ensuring that negation words only act on assertion relations within the boundary. For multiple parallel clinical results in the same sentence, evidence assertion nodes are generated separately, and each node references a common trial identifier and its own original text location identifier.

[0029] In a preferred embodiment, the version node includes a clinical event time node and a system effective time node. The graph processor connects the clinical event time nodes according to the order of events recorded in the clinical data to form a first directed version chain, and connects the system effective time node according to the order of effectiveness and ineffectiveness of the evidence assertion node in the clinical evidence knowledge graph to form a second directed version chain. Each evidence assertion node is connected to the event position in the first directed version chain and the effective interval in the second directed version chain.

[0030] In this embodiment, the clinical event time is taken from the registration, protocol change, data cutoff, result release, or review time point clearly recorded in the data. The graph processor converts the time point into a unified time representation and establishes a first directed version chain according to the order of events. When the data only records a date range, the start and end points of the range are saved and the time granularity is marked. When multiple events have the same recorded time point, parallel positions are established using the event order within the source without arbitrarily changing the clinical order. The system effective time is formed by the time when the evidence assertion is written into the activity graph, replaced by a subsequent assertion, or when participation in the activity query stops due to the withdrawal of the source. The previous node in the second directed version chain is connected to the next node through an effective inheritance edge, so that the order of clinical facts and the order of system recognition of facts are saved separately.

[0031] Specifically, each evidence assertion node is connected to an event location or event range on the first directed version chain, and to a valid interval defined by the effective start point and the invalidation end point on the second directed version chain. Evidence assertions that have not yet expired register the end point of the valid interval as open. When a subsequent version replaces the current assertion, the graph processor closes the valid interval of the current assertion with the system time when the new version enters the active graph, and establishes a new valid interval for the replacement assertion. The same clinical event time can correspond to multiple system effective versions, and the same system effective time point can also receive data belonging to different clinical event times. Based on this, the dual-chain structure supports querying by clinical history status and querying by data that the system has mastered at a certain point in time. Both types of queries obtain the source anchor point from the same evidence assertion node.

[0032] Please refer to Figure 2 In a preferred embodiment, the graph processor aligns the previous and subsequent versions of the source according to the source location and assertion semantic identifier, writes the semantic changes of the unaligned evidence content and the aligned evidence content into the version difference event; using the evidence assertion node corresponding to the version difference event as the seed node, it performs continuous directed traversal along the dependency relationship edges pointing to the downstream evidence association path until it reaches the path node that does not reference the changed version or the path node that does not have a downstream dependency relationship edge, forming an affected subgraph, and the affected subgraph is determined as the local reconstruction range.

[0033] Specifically, when aligning previous and subsequent versions, candidate fragment pairs are first formed using the chapter identifier and table header semantic path in the source anchor. Then, the normative subject reference, relational semantic identifier, and normative object reference of the evidence assertion corresponding to the candidate fragment are used to determine whether the semantics are consistent. For assertions where the paragraph order changes but the semantics remain consistent, only the source position mapping is updated. When there is no corresponding item for the candidate fragment, the addition or removal status is registered. When the candidate fragment can be matched but any semantic component changes, the semantic change status is registered. The graph processor generates semantic change markers according to the following formula.

[0034] In a preferred embodiment, when a source version update causes a change in paragraph order or table coordinates, the graph processor first queries the new version candidate segments using the canonical entity references and relational semantic identifiers of the segments corresponding to the old source anchors. Then, it compares the title semantic paths of the chapters in which the candidate segments are located and the references of adjacent segments. If the context of the same chapter and the semantic composition of the assertions are consistent, the new position is written to the current version source anchor while keeping the old anchor connected to the old version object. If there are multiple candidate positions, the positions are paired according to the relative order of adjacent assertions in the previous and subsequent versions. Assertions that still cannot be uniquely paired are not directly registered as semantic changes, but a position pending confirmation record is established and the relevant path is paused from entering the active index. The confirmed anchors then participate in the version difference judgment.

[0035] ; In the formula, This indicates the assertion of evidence in the preceding source version. This indicates the assertion of candidate evidence in the current source version. This indicates whether the two references to the same specification body are identical; the value is 1 if they are identical, but not 0 if they are both identical. This indicates whether the semantic identifiers of the two relationships are the same. This indicates whether the two references to the same normative object are identical. This indicates a semantic change marker. For example, if the drug entity, relation semantics, and result entity are all the same in the previous and current versions, all three judgment quantities are set to 1, the semantic change marker is set to 0, and the existing dependency edges are reused. If the current version replaces the result entity with the opposite state, then the object judgment quantity is set to 0, the semantic change marker is set to 1, and a version difference event pointing to the corresponding evidence assertion is generated. The downstream path traversal is initiated by the event.

[0036] Please refer to Figure 3 , Figure 3 The horizontal axis represents the pairing sequence number of the candidate evidence assertions in the preceding and following versions, and the vertical axis represents the strength of semantic change, with values ​​determined by semantic change markers. The results of smoothing the surrounding area are then displayed. The main curve corresponds to the change in assertion composition of the current source version relative to the previous source version. The verification line is used to indicate the medium change range that requires manual verification, and the reconstruction line is used to indicate the high change range that should generate version difference events and initiate downstream dependency propagation.

[0037] from Figure 3 It can be seen that the curves within the preceding sequence numbers mostly linger near the verification line, indicating that although some assertions have undergone changes in paragraph position, adjacent segments, or table coordinates, the main body of the specification still references them. Relational semantic identifiers and normative object reference Since no simultaneous destruction has occurred, the system can prioritize reusing existing dependency edges. The curve gradually crosses the reconstruction line and forms a continuous high-level region in the later stages, indicating that there are semantic differences in the candidate assertions regarding at least one of the subjects, relations, or objects. The value tends to change state, corresponding to the addition, removal, or semantic change of the record.

[0038] Preferably, the version difference event uses the source identifier, the previous version identifier, and the current version identifier as the event key. When the same source version is received repeatedly, the existing event is queried by pressing the event key. If it already exists and the status is complete, the original event and its propagation record are reused. If it exists but the status is interrupted, the uncompleted propagation queue position continues to be processed. Change assertions are not generated repeatedly or the valid interval is closed. The graph processor saves the assertion alignment result summary, change assertion set, and processing status in the event. The event key is recorded synchronously each time it is written to the propagation queue. After the local reconstruction is completed and the index is switched, the event status is changed to complete, so that the source repeated push and interruption recovery correspond to the same version difference processing record.

[0039] In a preferred embodiment, version difference events are recorded according to assertion addition, assertion removal, and assertion semantic change. The added record is connected to the current version assertion and establishes the system effective starting point for it. The removed record is connected to the previous version assertion and closes the original valid interval with the time when the current source version enters the graph. The semantic change record is connected to both the previous assertion and the current assertion and establishes a version substitution edge. The graph processor does not directly delete any historical nodes based on the change type, but converts the activity state of the previous assertion into a historical state and then starts dependency propagation with the current assertion. When the same evidence fragment contains both removed and added content, two change records are formed respectively and are classified into the same version difference event through a common event key.

[0040] In this embodiment, the construction of the affected subgraph uses a directed propagation queue to store the path nodes that are yet to be checked. All outgoing dependency edges of the seed node are written into the propagation queue. The graph processor reads the downstream path nodes connected by each edge and the versions they reference. If the downstream path references the changed version, the downstream path node and its continued association results are added to the affected subgraph. If the downstream path only shares a clinical entity with the change assertion but does not reference the changed version, the corresponding branch is stopped. During the traversal, a set of visited paths is established and duplicate visits are eliminated according to the path identifier. When a reference loop is encountered, the propagation direction formed by the first visit is retained until the propagation queue is empty, and the local reconstruction range consisting of the change assertion, the affected path and its boundary interface nodes is obtained.

[0041] Specifically, after the affected subgraph is generated, a local reconstruction list is formed. The list records the old path identifier, trigger assertion identifier, boundary interface identifier, and path type to be generated item by item. The graph processor rereads the currently effective assertions according to the query conditions of the old path, writes the nodes that still meet the query conditions into the candidate path, and retains the invalid assertions only in the old path reference. Each dependency edge of the new path is registered in reverse to the outgoing edge set of the trigger assertion. Items in the reconstruction list are marked as completed only after the candidate path, reverse dependency registration, and interface reference are written at the same time. Incomplete items are not included in the candidate index. When the reconstruction process stops due to an exception, the incomplete items continue to be processed according to the status of the list.

[0042] Furthermore, local reconstruction does not directly modify nodes and edges in the activity graph. The graph processor copies the boundary interface identifier of the affected subgraph to the candidate graph branch, regenerates the affected path based on the evidence assertion that is currently in the effective interval, connects the old path and the new path together to the version difference event and marks them as historical branch and candidate branch respectively. After the candidate branch completes the integrity verification, it switches the query entry point. The entry identifier, switching time and dependent source version of the original activity branch are written to the historical snapshot. After reading the snapshot entry point, the historical query only accesses the old path along the version dependency relationship saved in the snapshot, while the current query accesses the reconstructed path along the new activity entry point.

[0043] In a preferred embodiment, the graph processor forms a candidate assertion set based on the test identifier, and pairs the candidate assertions according to the standard entity references of the subject-object pair and the semantic identifiers of the assertion relationship; for evidence assertion nodes that match, their respective source anchors are retained and core reference edges are established between the evidence assertion nodes; for evidence assertion nodes that have semantic differences after pairing, mutually independent assertion branches are formed according to their respective connected version nodes.

[0044] Specifically, the candidate assertion set uses the trial identifier as the outer retrieval key and the canonical entity identifier of the assertion subject as the internal partition key. The graph processor compares the relation semantic identifier and the canonical entity identifier of the assertion object within the same partition. When all three correspond, it is determined to be a cross-source expression for the same clinical statement. Coreference relationship edges are established between evidence assertion nodes and the source type is written into the edge attribute. When candidate assertions come from registration data and research papers, the source anchor, clinical event time, and system effective time are still retained respectively. When querying the same statement, multiple sources are aggregated and displayed along the coreference relationship edge. When any source undergoes a version change, only the corresponding assertion branch is updated and the cross-source consistency status is rechecked through the coreference relationship edge.

[0045] In a preferred embodiment, candidate assertion pairing is processed in a stable order. Under the same test identifier, the system is first partitioned by normative subject entities, and then relationship buckets are established by relational semantic identifiers. Within the relationship buckets, the system is sorted by normative object entities and research context references. The graph processor sequentially retrieves the unpaired assertions from each source branch and establishes a pairing record only when the normative entity reference and the research context reference are consistent. An evidence assertion enters only one pairing record within the same source version. When duplicate expressions from the same source occur, the assertions are saved separately using the source position and connected by the duplicate expression edge. Cross-source pairing references representative assertions and incorporates the duplicate expression edge into the source tracing path to avoid duplicate text forming multiple core-reference branches with the same meaning.

[0046] Preferably, assertions with semantic differences after pairing are not merged through the overlay field. The graph processor establishes an independent assertion branch for each source version, and the root nodes of the branches are connected to the test identifier and the same specification subject entity. Branches with different relational semantics store their respective relational identifiers, and branches with different objects store their respective specification object references. Subsequent source versions only connect to the preceding branches along their own source revision dependencies. Core-referenced relation edges are only established between nodes with consistent semantic composition. When branches need to express substitution or conflict, corresponding relational nodes are created separately. This keeps the version chain within the source and the evidence relationship between sources separate from each other, avoiding the revision of one source from directly changing the state of the assertion of another source.

[0047] Furthermore, the same experiment may have registration entries, results announcements, and multiple research papers. The graph processor includes the data in the common candidate range according to the experiment identifier, and then limits the pairable assertions according to the research population identifier and the evaluation endpoint identifier. If the paper cites the registration identifier but reports a pre-set sub-population, then a limiting edge is established from the paper assertion to the sub-population entity and kept independent from the overall population assertion. If the results announcement and the paper use different wording for the same endpoint, then the endpoint entity and the result relationship are compared and standardized after terminology normalization. If the normalization results are consistent, a common reference relationship edge is established. If the normalization results are inconsistent, the branch is retained and the difference is handled by the conflict node.

[0048] In this embodiment, the research context reference consists of stable identifiers corresponding to the subject population entity, intervention entity, evaluation endpoint entity, and observation time point entity. The graph processor does not merge the identifiers into unparseable text fields, but instead establishes context relation edges around the evidence assertion nodes. When pairing across sources, the canonical references of the context entities are compared edge by edge. When there is a complete correspondence, the establishment of coreference relation edges is allowed. When there is only a partial correspondence, candidate association edges are established and the non-corresponding context types are recorded. Candidate association edges do not participate in conflict judgment and evidence path aggregation. After the missing context is supplemented by subsequent sources, the pairing is re-executed and the candidate association edges that meet the conditions are replaced with coreference relation edges.

[0049] Furthermore, later-acquired data may record experimental facts of clinical events earlier than the latest clinical events in the current atlas. The atlas processor inserts the corresponding event node into the actual time position of the first directed version chain based on the clinical event time in the data. At the same time, it establishes a new system effective node at the end of the second directed version chain based on the time the data enters the active atlas. The position of the clinical event asserted by the original evidence does not move due to the late arrival of the data. If the late data corrects the earlier event, the new assertion connects the position of the earlier clinical event and connects to the current system effective interval. Historical system time point queries will not read late assertions, but assertions can be read when the current system time point is combined with early clinical event queries, thereby maintaining the independence of the order of event occurrence and the order of data arrival.

[0050] In a preferred embodiment, each system effective time node in the second directed version chain is connected to the valid start or valid end point of the corresponding evidence assertion node. The graph processor receives a retrieval request containing the clinical event query time point and the system effective query time point. First, it filters evidence assertion nodes that are no later than the clinical event query time point according to the first directed version chain. Then, it retains evidence assertion nodes whose valid intervals cover the system effective query time point according to the second directed version chain. The retained evidence assertion nodes generate a query subgraph.

[0051] In this embodiment, the clinical event query time point in the retrieval request limits the scope of clinical facts to be observed, and the system effective query time point limits the evidence version that the system can use at the corresponding time. The graph processor performs dual time determination on each candidate evidence assertion. Only nodes whose clinical event time is not later than the clinical event query time point and whose system effective interval covers the system effective query time point are included in the query subgraph. The determination process is expressed by the following formula.

[0052] Specifically, before the retrieval request enters the dual-time determination, the atlas processor converts the date, date range, and open time point into a representation consistent with the time node. When the clinical event query time point is defaulted, the position of the latest event in the active atlas is taken. When the system effective query time point is defaulted, the switching time corresponding to the active index is taken. When the query time point falls within the date range only recorded in the data, the event node is selected based on the range inclusion relationship without presuming the specific time. When the query time point is earlier than the starting point of the event range of the evidence assertion, the assertion is excluded. When the query time point is later than the end point of the event range, the system effective range is combined for judgment. All conversion results are saved to the retrieval request record and returned with the query subgraph.

[0053] ; In the formula, e represents the evidence assertion node to be judged. Indicates the time point for clinical event query. Indicates the point in time when the system query takes effect. This indicates the time of the clinical event corresponding to the evidence assertion. This indicates the starting point for the effectiveness of the evidentiary assertion in the system. This indicates the endpoint of failure. The indicator value is 1 when the condition within the square brackets is true and 0 when it is false. When the value is 1, the node is retained; when the value is 0, the node is excluded. For example, if the clinical event time of a certain assertion is March 1, 2026, and the effective period is from March 5, 2026 to April 1, 2026, when the clinical event query time is March 10, 2026 and the system effective query time is March 20, 2026, both indicator values ​​are 1 and the assertion is written into the query subgraph. When the system effective query time is changed to April 2, 2026, the second indicator value is 0 and the assertion is excluded from the corresponding query subgraph.

[0054] Please refer to Figure 4 , Figure 4 The horizontal axis represents the system retrieval date, and the vertical axis represents the number of evidence assertions involved in the dual-time determination on that date; the solid line represents... To preserve the assertion size, dashed lines represent assertion sizes that are excluded because the clinical event time or system validity interval does not meet the criteria.

[0055] Figure 4 around and The joint query process unfolds, in which... The determination of whether an assertion falls within the scope of specified clinical facts and The common constraint is whether the system has received and still acknowledges the availability of the assertion at the corresponding point in time. The rapid increase in the number of assertions retained at the front of the curve indicates that as registration data, protocol revision data, and outcome data are successively written into the activity graph, more and more assertions simultaneously meet the two conditions: the clinical event is no later than the query time and the system's effective period covers the query time.

[0056] As the date progresses into the later stages, the retention curve tends to level off while the exclusion curve rises slowly, reflecting that the system retains currently searchable evidence without mixing withdrawn, substituted, or future clinical events into the current query subgraph. Figure 4 Emphasizing the independent role of the left-closed / right-open systemic interval and the clinical event chain: the same assertion only applies when... and The query results are only included when both dimensions meet the formula conditions; otherwise, they are only saved as traceable objects in historical snapshots or candidate branches.

[0057] Preferably, the system effective interval adopts a left-closed and right-open time boundary. Evidence assertions can enter the query subgraph at the time corresponding to the effective start point, and will not enter the current query subgraph at the time corresponding to the failure end point. When the failure end point is open, the graph processor determines the second indicative condition as the query time point is not earlier than the effective start point. When the clinical event only has a date range, the event range end point is compared with the clinical event query time point and the time granularity identifier is retained in the query subgraph. When multiple evidence assertions take effect in the same switching transaction, they share the system effective time node but save the effective interval edge separately. The query record can thus replay the inclusion status of each assertion at the same time.

[0058] Specifically, the query subgraph is based on the evidence assertion nodes that satisfy the dual-time determination. It connects the drug entity and the clinical research entity along the entity reference edge, connects the original data location along the source anchor point, connects other effective sources of the same clinical statement along the coreference relationship edge, and reads the evidence association path supported by the assertion along the dependency relationship edge. If the historical snapshot entry corresponds to the system effective query time point in the retrieval request, the path index at that time is read from the historical snapshot without calling the subsequent reconstructed path. When determining the open valid interval, cases where the query time point is later than the effective start point are considered as overwrites. The query result saves both the clinical event query time point and the system effective query time point so that the same set of nodes can be reproduced when read again.

[0059] In a preferred embodiment, the dependency edge associates the version identifier of the upstream evidence assertion node with the path identifier of the downstream evidence association path; the graph processor reads the changed version identifier from the version difference event, generates a dependency signature for each evidence association path consisting of the dependent version identifier, writes the path node whose dependency signature contains the changed version identifier into the propagation queue, continues to traverse the downstream dependency edge according to the propagation queue, and terminates the corresponding traversal branch when the dependency signature does not contain the changed version identifier, while recording the visited path identifier.

[0060] Specifically, the dependency signature is obtained by concatenating the version identifiers directly referenced by the evidence association path according to the stable sorting rules and then calculating the digest. The stable sorting is based on the source identifier as the main order and the evidence assertion identifier as the secondary order. The same dependency set generates the same signature accordingly. In addition to storing the signature, the path node also stores a list of version identifiers that constitute the signature. The version difference event carries the previous version identifier and the current version identifier. When the propagation begins, the graph processor uses the list of version identifiers to determine whether the path directly references the previous version. If it does, the path is written into the propagation queue and the candidate dependency signature is recalculated. When the path is derived from other paths, the upstream path identifier and the upstream signature are written into the downstream dependency record at the same time, so that the source version change can continue to propagate along the inter-path dependencies.

[0061] In this embodiment, the propagation queue adopts a first-in-first-out ordered record structure. Each record contains the identifier of the path to be checked, the trigger version identifier, and the upstream path identifier. After the graph processor retrieves the record, it queries the outgoing dependency edges of the corresponding path, writes the downstream paths that have not yet been visited into the queue, and adds the current path identifier to the visited set. If the version identifier list of the downstream path does not contain the trigger version and its upstream signature has not changed, it will not be written into the queue. If multiple propagation branches reach the same path, the duplicate records are merged using the visited set and all trigger sources are saved. When there is a circular reference, the propagation level when it first enters the queue is used as the processing position. The circular records that arrive later only supplement the trigger source and do not expand the downstream again, so that the affected range is traversed within a limited set of paths.

[0062] Furthermore, during path reconstruction, the graph processor reads all trigger sources recorded in the propagation queue, reassembles the dependency version list for evidence assertions still in the effective range, recalculates the dependency signatures and compares them with the old signatures. If the signatures are the same, the original path content is retained and the inspection record is updated. If the signatures are different, a new path node is generated in the candidate graph branch. The new path node connects to the currently effective version and connects to the old path node through the version replacement edge. Multiple new paths triggered by the same version difference event share the event identifier. Subsequently, all paths and their boundary nodes involved in this local reconstruction can be retrieved from the event identifier.

[0063] In a preferred embodiment, the graph processor retains interface nodes at the connection between affected and unaffected subgraphs, writes the evidence assertion nodes and dependency edges generated by local reconstruction into a candidate graph branch that is separate from the current graph, performs referential integrity verification on the candidate graph branch and interface nodes, and switches the query index from the current graph branch to the candidate graph branch after the verification passes, and connects the current graph branch before the switch to the historical snapshot.

[0064] Specifically, interface nodes are clinical entity nodes, trial nodes, or source nodes that are jointly referenced by affected and unaffected subgraphs and whose content has not changed. Candidate graph branches only copy the stable identifier of the interface node and save the reference to the read-only interface node in the active graph. They do not copy the complete content of the unaffected subgraph. The assertion nodes and path nodes obtained from local reconstruction are written to the candidate branch with new version identifiers. The start and end points of dependency edges must fall on the candidate branch node or the registered interface node. The graph processor builds an independent query index on the candidate branch. The active query still reads the current graph index before switching. The writing, verification, and index construction of the candidate branch will not change the current graph entry.

[0065] Preferably, the reference integrity check examines each evidence assertion in the candidate branch to see if it has a readable source anchor, whether the version node is connected to both time axes, whether the two ends of the dependency edge exist, and whether the dependency version list of the path node is consistent with the actual incoming edge. The check record saves the check object identifier and check status. After all check statuses are consistent, an index switching record is generated. The switching record includes the current index identifier, candidate index identifier, candidate branch version identifier, switching time, and historical snapshot entry. The graph processor updates the active index pointer and writes it to the switching record in the same storage transaction. The original active index pointer is maintained while the transaction is not completed. After the transaction is completed, new retrieval requests read the candidate index, and retrieval requests that have already started continue to read the index version locked when it was created.

[0066] In a preferred embodiment, the graph processor performs object semantic consistency verification on the evidence assertion nodes connected by coreference edges. When there is a mutually exclusive assertion relationship between evidence assertion nodes that have the same subject-object standard entity references, a conflict node is established connected to the corresponding evidence assertion node. The conflict node is associated with the source anchor and version node of the corresponding evidence assertion node, and the evidence association path referencing the evidence assertion node is connected through dependency edges.

[0067] Specifically, the object semantic consistency check compares the standardized object entities and their state directions within the coreference candidate range. The graph processor uses the relational semantic table to store mutually exclusive relation pairs. For example, confirmation relations and denial relations, occurrence relations and non-occurrence relations have opposite directions. A conflict node is established only when the subject and object standard entities reference the same thing and the relational semantic table determines that the directions are mutually exclusive. Assertions with different research populations, evaluation endpoints, or observation points maintain independent contexts and do not enter the same conflict node. The conflict node stores the set of assertion identifiers participating in the conflict and connects the source anchor point and version node of each assertion. The evidence association path that references any conflict assertion adds a dependency edge pointing to the conflict node. The path query obtains the conflict content and their respective sources simultaneously.

[0068] Furthermore, conflict nodes form version branches as the version status of the participating assertions is updated. When a source releases revised content and invalidates the original mutually exclusive assertion, the version difference event is mapped to the original assertion and reaches the conflict node along the dependency edge. The graph processor re-executes the object semantic consistency check in the candidate graph branch. When the currently effective assertions are no longer mutually exclusive, the system effective range of the old conflict node is closed, and a conflict-free coreference path is established. If the mutual exclusion relationship still exists, a new conflict node connecting the current version assertion is generated. The old conflict node and its source anchor point are retained in the historical snapshot. When searching by the old system effective query time, the conflict state at that time can still be restored.

[0069] In a preferred embodiment, the graph processor reads the dependency edges of each reconstructed evidence association path within the candidate graph branch, verifies whether the connected evidence assertion nodes have source anchors, and verifies that the corresponding version node is within the valid range of the selected query time point based on the clinical event time and the system effective time. When the verification results are consistent, a switching record containing the current query index identifier and the candidate query index identifier is generated, the active index pointer is updated according to the switching record, and the active index pointer before the switch is written to the entry node of the historical snapshot.

[0070] In this embodiment, the graph processor reads all incoming dependency edges for each reconstructed evidence association path, and checks the evidence assertion identifiers connected to the edges against the dependency version list stored in the path item by item. If an evidence assertion lacks a source anchor, the source anchor cannot locate the original object, or the version node is not connected to the dual time axis, it is registered as a verification inconsistency. If the version node exists, dual time determination is performed based on the query time point corresponding to the path. Assertions not in the selected valid interval are not allowed to be written to the current path. Paths with consistent verification generate path check summaries and write them to the candidate index. After all candidate paths have completed verification, a switching record is created. The update of the active index pointer and the writing of the historical snapshot entry are committed by the same transaction, so that there is a complete index entry that can be read at any time.

[0071] Specifically, the switch record serves as a connection between the current graph version and historical graph versions. It records the index before the switch, the index after the switch, the triggering version difference event, and a summary of candidate branch checks. After the active index pointer is updated, the original index pointer is written to the historical snapshot entry and connected to the previous historical snapshot, forming a snapshot chain arranged according to the system effective time. The recovery operation reads the switch record with the most recent completion status and points the active index pointer to the candidate index in the record. If the switch transaction has not yet reached a completion status, the record is ignored and the original index continues to be used. The candidate branch is retained as a pending object and does not participate in the retrieval. After re-verification, the switch can continue or the graph processor can revoke its temporary index.

[0072] In one embodiment, the system receives registration data, protocol revision data, and results paper from a clinical trial of a new drug. The graph processor establishes a common clinical research entity based on the trial identifier, converts statements about the primary endpoint in the registration data into first evidence assertions, converts modifications to the endpoint in the protocol revision data into second evidence assertions, and connects the second evidence assertions to the first evidence assertions through source revision dependencies. The statements in the results paper reporting the modified endpoint results form third evidence assertions. The third evidence assertions are connected to the second evidence assertions through coreference edges, thus preserving the source anchor of the paper. The three assertions are respectively connected to their clinical event time and system effective interval. When querying the original protocol time, the system reads the first evidence assertions, and when querying the status after the revision takes effect, it reads the second evidence assertions and the third evidence assertions that they co-refer to.

[0073] In another embodiment, regulatory review materials correct the conclusions in existing results papers. The graph processor registers the new review materials as subsequent source versions, aligns the corrected statements with the original review assertions through chapter position and semantic composition, and generates version difference events when the semantics of the objects before and after the correction are different. The version difference events point to the original assertion and the current assertion and carry the changed version identifier. The propagation queue retrieves the drug-indication association paths, evidence summary paths, and conflict nodes that cite the original assertion, incorporates the paths that actually cite the original version into the affected subgraph, reconstructs the paths of candidate graph branches with the currently effective assertions, switches the active index after verification by source anchors and dual time effective intervals, and saves the original association paths and original conflict states to historical snapshots.

[0074] In one embodiment, when the data publisher withdraws the entire result document, the graph processor registers the withdrawal notification as a new source version and links the withdrawn document identifier in the notification to the original source object. It establishes a removal change record for each active evidence assertion generated by the original source object, closes the corresponding system effective interval based on the time the withdrawal notification enters the graph, and the propagation queue only includes evidence association paths of directly cited assertion versions. Other source assertions connected by co-reference edges maintain their original effective intervals. Candidate graph branches are regenerated to have paths still supported by effective sources, and paths supported only by withdrawn sources are closed. Historical snapshots continue to retain evidence nodes, conflict nodes, and source anchors before the withdrawal.

[0075] In a preferred embodiment, when multiple retrieval requests and source updates arrive simultaneously, each retrieval request obtains the active index pointer and its graph version identifier before starting to read, and reads the corresponding version at a fixed time during the duration of the request. The source update is always written to the candidate graph branch and does not modify the locked active index. New retrieval requests arriving after the switch record is submitted read the new active index. Retrieval requests that started before submission are completed release the original version reference. The evidence storage converts duplicate index pages into shared read-only pages only after the original version no longer has an active read reference and its entry has been written to the historical snapshot. The historical snapshot still retains reachable references to nodes, edges, and source objects, avoiding the generation of query subgraphs that mix some new paths and some old paths during the update process.

[0076] In this embodiment, when original data parsing fails, source identifier conflicts occur, or candidate branch verification is inconsistent, the graph processor writes the corresponding source record, evidence fragment, or path node into the isolation state and saves the failure position. The isolated object does not enter the active graph but retains the source reference with the original object. After format repair or entity verification is completed, parsing and assertion construction are re-executed from the failure position. Existing standardized entity nodes and unaffected paths are directly reused. The regenerated evidence assertion connects the two time axes according to the source version and re-enters the difference propagation process. The candidate branch only obtains the active index entry after both the reference integrity and the effective time interval of the two time are verified. Historical snapshots and existing active indexes remain read-only during the isolation process.

[0077] Furthermore, the evidence storage stores version references for node records, relation edge records, source objects, and index records respectively. When the graph processor deletes a candidate branch, it only removes the temporary index reference to the candidate node, without deleting assertion nodes and source objects that are still reachable from historical snapshots. The content summary value of the source object is used to verify that the read content is consistent with the content when the source anchor point was established. The graph backup simultaneously saves the active index pointer, the switching record chain, and the historical snapshot entry. After restoration, it first verifies the status of the switching record completion, and then rebuilds the entity index, source index, and version index that are reachable from the entry node, so that the current query, dual-time historical query, and version difference tracking continue to use the same set of stable node identifiers.

[0078] In a complete operational scenario, after clinical data from multiple sources enters the system, source records, standardized clinical entity citations, and evidence assertion nodes are formed sequentially. Evidence assertions are connected to the original text through source anchors and enter the dual time axis through version nodes. Cross-source identical assertions maintain independent tracing using coreference edges. Mutually exclusive assertions save their respective versions using conflict nodes. Evidence association paths record the assertion versions that are directly dependent and form dependency signatures. When the source version changes, version difference events are generated by semantic change markers. The propagation queue determines the affected subgraph based on the dependency signature. The local reconstruction results are written into the candidate graph branch and undergo source anchor, dual time interval, and dependency integrity verification. After the index switching record is submitted, the activity entry is updated and the original entry is written into the historical snapshot. The current search and the historical search read the complete evidence subgraph along the corresponding entry.

Claims

1. A knowledge graph-based system for analyzing clinical evidence association in new drug trials, characterized in that, It includes a graph processor and an evidence storage device connected to the graph processor, the evidence storage device storing a clinical evidence knowledge graph; The graph processor converts multi-source clinical evidence into evidence assertion nodes, associates each evidence assertion node with a clinical entity node and its source location, and connects version nodes that record the clinical event time and the system effective time. It also establishes dependency edges between evidence assertion nodes based on source revision dependencies and conclusion dependencies. The graph processor compares the previous and subsequent versions of the source to generate version difference events, maps the version difference events to the corresponding evidence assertion nodes, determines the affected evidence association paths along the dependency edges, performs local reconstruction of the evidence association paths, and saves the historical snapshot before the update.

2. The knowledge graph-based new drug clinical evidence association analysis system according to claim 1, characterized in that, The graph processor segments multi-source clinical evidence according to document paragraph boundaries and table data boundaries, and converts the segmented evidence fragments into evidence assertion nodes composed of subject-object pairs and assertion relations that limit the semantics of subject-object pairs. The clinical entity nodes include drug entity nodes and clinical research entity nodes. The subject-object pair references the clinical entity nodes. The evidence assertion nodes are bound to source anchors consisting of source identifiers and original text location identifiers, and are associated with different sources of evidence from the same clinical trial through trial identifiers.

3. The knowledge graph-based new drug clinical evidence association analysis system according to claim 1, characterized in that, The version nodes include clinical event time nodes and system effective time nodes. The graph processor connects the clinical event time nodes according to the order of events recorded in the clinical data to form a first directed version chain, and connects the system effective time nodes according to the order of effectiveness and ineffectiveness of the evidence assertion nodes in the clinical evidence knowledge graph to form a second directed version chain. Each evidence assertion node is connected to the event position in the first directed version chain and the valid interval in the second directed version chain.

4. The knowledge graph-based new drug clinical evidence association analysis system according to claim 1, characterized in that, The graph processor aligns the previous and subsequent versions of the source according to the source location and assertion semantic identifier, and writes the semantic changes of the unaligned evidence content and the aligned evidence content into the version difference event. Using the evidence assertion node corresponding to the version difference event as the seed node, a continuous directed traversal is performed along the dependency edges pointing to the downstream evidence association path until a path node that does not reference the changed version or does not have a downstream dependency edge is reached, forming an affected subgraph, and the affected subgraph is determined as the local reconstruction range.

5. A knowledge graph-based new drug clinical evidence association analysis system according to claim 2, characterized in that, The graph processor forms a candidate assertion set based on the test identifier, and pairs the candidate assertions according to the standard entity references of the subject-object pair and the semantic identifiers of the assertion relationship; For paired evidence assertion nodes, their respective source anchors are preserved and a coreference edge is established between the evidence assertion nodes. For evidence assertion nodes that have semantic differences after pairing, independent assertion branches are formed according to the version nodes they are connected to.

6. The knowledge graph-based new drug clinical evidence association analysis system according to claim 3, characterized in that, Each system effective time node in the second directed version chain is connected to the valid start or valid end point of the corresponding evidence assertion node. The graph processor receives a retrieval request containing the clinical event query time point and the system effective query time point. First, it filters evidence assertion nodes that are no later than the clinical event query time point according to the first directed version chain. Then, it retains evidence assertion nodes whose valid intervals cover the system effective query time point according to the second directed version chain. Finally, it generates a query subgraph from the retained evidence assertion nodes.

7. A knowledge graph-based new drug clinical evidence association analysis system according to claim 4, characterized in that, The dependency relationship edge associates the version identifier of the upstream evidence assertion node with the path identifier of the downstream evidence association path; The graph processor reads the changed version identifier from the version difference event, generates a dependency signature consisting of a dependency version identifier for each evidence association path, writes the path node whose dependency signature contains the changed version identifier into the propagation queue, continues to traverse the downstream dependency relationship edges according to the propagation queue, and terminates the corresponding traversal branch when the dependency signature does not contain the changed version identifier, while recording the visited path identifiers.

8. A knowledge graph-based new drug clinical evidence association analysis system according to claim 4, characterized in that, The graph processor retains interface nodes at the connection between the affected subgraph and the unaffected subgraph, and writes the evidence assertion nodes and dependency edges generated by local reconstruction into the candidate graph branch that is separated from the current graph. Perform referential integrity verification on the candidate graph branch and the interface node, and after the verification passes, switch the query index from the current graph branch to the candidate graph branch, and connect the current graph branch before the switch to the historical snapshot.

9. A knowledge graph-based new drug clinical evidence association analysis system according to claim 5, characterized in that, The graph processor performs object semantic consistency verification on the evidence assertion nodes connected by the coreference relationship edges. When there is a mutually exclusive assertion relationship between evidence assertion nodes with the same subject-object standard entity references, a conflict node is established connected to the corresponding evidence assertion node. The conflict nodes are respectively associated with the source anchor point and version node of the corresponding evidence assertion node, and are connected to the evidence association path referencing the evidence assertion node through the dependency edge.

10. A knowledge graph-based clinical evidence association analysis system for new drugs according to claim 8, characterized in that, The graph processor reads the dependency edges of each reconstructed evidence association path within the candidate graph branch, verifies whether the connected evidence assertion nodes have source anchors, and verifies that the corresponding version node is within the valid range of the selected query time based on the clinical event time and the system effective time. When the verification results are consistent, a switch record containing the current query index identifier and the candidate query index identifier is generated. The active index pointer is updated according to the switch record, and the active index pointer before the switch is written to the entry node of the historical snapshot.

Images