A knowledge graph-based drug clinical trial risk traceability management method

Abstract

The application discloses a kind of based on knowledge graph's drug clinical test risk traceability management method, comprising the following steps: based on the multi-source heterogeneous data of drug clinical test constructs entity domain hypergraph and relationship domain hypergraph;Using cross-domain interactive fusion mechanism, realize the real-time mutual feedback update of entity domain hypergraph node embedding and relationship domain hypergraph hyperedge embedding;According to stable two-domain embedding information generates risk evolution propagation path;Using recursive causal link back-propagation algorithm extracts risk event end-to-end data relationship, realizes drug clinical test risk traceability and audit tracking.The application realizes the time sequencing of drug clinical test risk, computable traceability.

Description

Technical Field

[0001] This invention relates to the field of pharmaceutical information technology, and in particular to a method for risk traceability management of drug clinical trials based on knowledge graphs. Background Technology

[0002] Drug clinical trial risk tracing management technology involves the identification, monitoring, and causal link analysis of risk events during the clinical trial process, and is one of the core technologies for drug clinical trial quality control. Drug clinical trials typically involve numerous entities, including subjects, research centers, clinical trial institutions, drug formulations, and trial visit points. The interactions between these entities constitute the multi-source heterogeneous characteristics of trial data. Existing drug clinical trial risk management methods mainly rely on single risk event detection and isolated risk analysis, using statistical analysis or rule matching to identify and warn of risk events. However, due to the failure to effectively establish clear causal relationships between various risk events, accurate tracing and tracking of risk events are difficult, failing to meet the precision requirements of risk monitoring, quality auditing, and compliance analysis.

[0003] In recent years, knowledge graph technology has been gradually applied to risk analysis in drug clinical trials. It uses a structured graph representation to represent trial entities and risk events as nodes, connecting them through simple relationships to describe data associations. However, existing knowledge graph methods are mainly limited to single-layer graph representations, making it difficult to describe the complex high-order interactions in clinical trials and the temporal propagation paths of risk events. They fail to effectively address the cross-domain feature interaction and fusion problem between entity relationships and risk event propagation paths. Existing methods cannot form a clear temporally directed risk propagation chain under multi-source heterogeneous data conditions, thus failing to accurately achieve recursive causal tracing of risk events from back to front, reducing the effectiveness, accuracy, and practicality of risk tracing.

[0004] Therefore, how to provide a knowledge graph-based method for risk tracing and management in drug clinical trials 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 knowledge graph-based method for risk tracing and management in drug clinical trials. Addressing the challenges of risk tracing caused by the lack of unified, inferable semantic representations in multi-source, heterogeneous clinical trial data and the inability to construct temporally directed causal chains, this invention proposes a composite dual-hypergraph recurrent neural network. This network constructs entity-domain hypergraphs and relation-domain hypergraphs, combined with a cross-domain interaction fusion mechanism, to achieve temporal modeling of clinical trial risk events and calculation of causal links. This invention realizes temporal and computable tracing of drug clinical trial risks.

[0006] A drug clinical trial risk tracing management method based on knowledge graph according to an embodiment of the present invention includes: Based on multi-source heterogeneous data from drug clinical trials, an entity domain hypergraph is constructed, with trial subjects and trial elements as nodes and the interaction relationships between trial subjects and trial elements as hyperedges. Based on multi-source heterogeneous data, risk events that occur during the experiment are identified, and a relational domain hypergraph is constructed with risk events as nodes and risk propagation relationships between risk events as hyperedges. The entity domain hypergraph and the relation domain hypergraph are simultaneously input into the composite dual hypergraph recurrent neural network. Through the cross-domain interaction fusion mechanism, the node embedding of the entity domain hypergraph is updated in real time according to the hyperedge embedding of the relation domain hypergraph. Based on the node embedding of the updated entity domain hypergraph, the hyperedge embedding of the corrected relation domain hypergraph is performed in real time. The cross-domain interaction and fusion mechanism is executed in a loop, so that the node embedding of the entity domain hypergraph and the hyperedge embedding of the relation domain hypergraph feed back to each other in real time in each iteration until the embedding information of the two domain hypergraphs reaches stability. Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, a risk evolution and propagation path is generated. A recursive causal link back-inference algorithm is used to extract end-to-end data relationships of risk events from the risk evolution and propagation path, and to conduct risk tracing and audit tracking in drug clinical trials.

[0007] Optionally, the construction of the entity domain hypergraph specifically involves: The trial subjects and elements are extracted from multi-source heterogeneous data of drug clinical trials. The trial subjects include subjects, research centers, and clinical trial institutions. The trial elements include drug formulations, trial visit points, and trial operation procedures. The interaction relationship between each experimental subject and its corresponding experimental element is mapped as a hyperedge, and the set of nodes connected by the multiple experimental subjects and experimental elements associated with the interaction relationship is used as the hyperedge. Weighted superedges are obtained by assigning weights to the superedges based on the frequency, duration, and intensity of the interaction. The experimental subjects and elements are represented as node embeddings in a unified semantic space according to their respective experimental stages, timestamps, and associated weighted hyperedges, resulting in an entity domain hypergraph.

[0008] Optionally, the construction of the relational domain hypergraph specifically involves: Based on multi-source heterogeneous data from drug clinical trials, risk events that occur during the trial are identified, including abnormal data events, process deviation events, and events where key quality indicators exceed limits. Based on the order of occurrence and correlation of risk events, determine the propagation path between risk events, and associate risk events with direct causal propagation relationships; Each risk event is mapped to a node in the relational domain hypergraph, and the propagation path consisting of multiple risk events with direct causal propagation relationships is mapped to a hyperedge, forming a set of risk event nodes; Based on the correlation strength, propagation speed and impact on the overall risk of the experiment among risk events, the hyperedges of the relational domain hypergraph are weighted to obtain weighted hyperedges. Based on the experimental phase, timestamp, and corresponding weighted hyperedge information at the time of the risk event, it is represented as node embedding in a unified semantic space, resulting in a relational domain hypergraph.

[0009] Optionally, the composite dual-hypergraph recurrent neural network includes an input embedding layer, an intra-domain message passing layer, a cross-domain interaction fusion layer, an attention fusion layer, a gated recurrent update layer, and a temporal constraint and convergence output layer. The input embedding layer is used to receive node and hyperedge input data from the entity domain hypergraph and the relation domain hypergraph respectively. It generates corresponding time position codes based on the test stage and timestamp, and then fuses them with the initial features of the nodes and hyperedges to generate an initial embedding with time information. The intra-domain message passing layer is used to receive the initial embedding output from the input embedding layer. It performs bidirectional message passing of intra-domain features within the entity domain hypergraph and the relation domain hypergraph, respectively, and generates entity domain node representations and relation domain hyperedge representations. The cross-domain interaction fusion layer is used to receive entity domain node representations and relation domain hyperedge representations output by the intra-domain message passing layer. Through the cross-domain interaction fusion mechanism, it performs bidirectional interaction of cross-domain features to generate comprehensive feature representations. The attention fusion layer is used to receive the comprehensive feature representation of the cross-domain interaction fusion layer. It assigns different attention weights to the intra-domain local features and cross-domain interaction features of the entity domain and the relation domain respectively through the attention mechanism, and generates fused features by weighting and fusing according to the size of the attention weights. The gated loop update layer is used to receive the fusion features output by the attention fusion layer, and performs gated loop updates on the node feature embeddings of the entity domain and the hyperedge feature embeddings of the relation domain respectively, forming the updated embedding representation at the current time. The temporal constraint and convergence output layer is used to apply temporal directional constraints and topological consistency constraints to the embedded representations output by the gated cyclic update layer. It evaluates the magnitude of feature embedding changes after two consecutive cyclic updates. When the magnitude of the embedding changes reaches within a preset convergence threshold, the network is determined to be converged, and stable embedding representations of entity domain nodes and relation domain hyperedges are output.

[0010] Optionally, the cross-domain interaction and fusion mechanism specifically includes: Based on the entity domain node representation and relation domain hyperedge representation output by the intra-domain message passing layer, cross-domain message paths from relation domain hyperedges to entity domain nodes and from entity domain nodes to relation domain hyperedges are constructed respectively. In the cross-domain message path from relational hyperedge to entity node, the relational hyperedge representation is mapped to a query vector, and the entity node representation is mapped to a key vector and a value vector. The attention weight of the relational hyperedge representation to the entity node representation is calculated through a multi-head self-attention mechanism, and the entity node representation is weighted and fused with the attention weight to form the cross-domain fusion feature of the entity node. In the cross-domain message path from entity domain nodes to relation domain hyperedges, entity domain node representations are mapped to query vectors, and relation domain hyperedge representations are mapped to key vectors and value vectors. The attention weights of entity domain node representations to relation domain hyperedge representations are calculated through a multi-head self-attention mechanism, and the relation domain hyperedge representations are weighted and fused using the attention weights to form the cross-domain fusion features of relation domain hyperedges. The cross-domain fusion features of entity domain nodes and the cross-domain fusion features of relation domain hyperedges obtained through bidirectional cross-domain message channels are merged into a comprehensive feature representation of cross-domain interaction.

[0011] Optionally, the hyperedge embedding of the corrected relation domain hypergraph specifically includes: Based on the node embeddings of the updated entity domain hypergraph, calculate the change in node embeddings of the risk event node set corresponding to each relation domain hyperedge; Based on the change in node embedding, calculate the adjustment weight of the current hyperedge embedding for each relation domain hyperedge; Based on the adjusted weights, the hyperedge embeddings of the relation domain hypergraph are weighted and fused to obtain the real-time corrected hyperedge embeddings of the relation domain. Based on the actual occurrence order of the risk event node set corresponding to the hyperedge embedding of the relation domain after real-time correction, a time-directed constraint is applied; Perform topological consistency verification on hyperedge embeddings of relation domains subject to time-directed constraints; The historical hyperedge embedding of the relation domain hypergraph is replaced by the hyperedge embedding of the relation domain hypergraph after the hyperedge embedding has been verified by time-directedness constraints and topological consistency, so as to complete the real-time correction of the hyperedge embedding of the relation domain hypergraph.

[0012] Optionally, the generation of the risk evolution and propagation path specifically includes: Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, calculate the similarity between the embedding features of each hyperedge in the relation domain hypergraph and the embedding features of the corresponding risk event node set. Based on the calculated similarity, all relation domain hyperedges are sorted and filtered, and hyperedges exceeding the preset similarity threshold are extracted as candidate risk propagation paths; For the candidate risk propagation paths, the candidate risk propagation paths are arranged chronologically according to the timestamps corresponding to each node in the risk event node set and the actual occurrence order of the risk events. The confidence of each candidate risk propagation path is calculated by weighting the path confidence of the hyperedge similarity in the arranged candidate risk propagation paths. The confidence level of each path in the candidate risk propagation path is filtered according to the preset path confidence threshold, and the paths with confidence levels higher than the preset threshold are retained to obtain the risk evolution and propagation path. The risk evolution and propagation path is represented by sequentially associating risk event nodes with their corresponding timestamps, thus generating a time-series risk evolution and propagation path.

[0013] Optionally, the recursive causal link back-reasoning algorithm is specifically as follows: Based on the risk evolution and propagation path, select the target risk event nodes to be traced and establish an initial recursive node set; Based on the target risk event node in the initial recursive node set, determine the predecessor risk event node that is directly related to the target risk event node and has an earlier timestamp, and establish the predecessor node set for the current recursive level. Based on the node embedding features of the predecessor node set at the current recursive level, calculate the cosine similarity between the node embedding features of each node in the predecessor node set and the target risk event node embedding features, and determine the nodes with a cosine similarity higher than the preset cosine similarity threshold as valid predecessor nodes. For valid precursor nodes, based on the embedded features and corresponding timestamps, we recursively backtrack to earlier moments in the risk evolution and propagation path, identifying upstream precursor risk event nodes layer by layer, and using the similarity threshold of node embedded features for layer-by-layer screening during the recursive backtracking process. During the recursive reverse process, topological consistency is verified for each recursive path, and recursion is terminated in a timely manner for paths that do not meet the topological consistency requirements. During the recursive backward path formation process, the path confidence is calculated by multiplying the embedding feature similarity of consecutive nodes in each recursive path, and only backward paths with a path confidence threshold are retained. The nodes in the retained reverse path are arranged in chronological order according to their respective timestamps to obtain a recursive causal link with chronological order. The above recursive backward process is repeated for each target risk event node until the end-to-end data relationship link of all target risk event nodes in drug clinical trial risk tracing is completed, and the recursive causal link of all risk events is output for drug clinical trial risk tracing and audit tracking.

[0014] The beneficial effects of this invention are: (1) By constructing entity domain hypergraphs and relation domain hypergraphs and adopting cross-domain interaction fusion mechanism, this invention realizes unified semantic representation of multi-source heterogeneous clinical trial data and real-time mutual feedback update of dual-domain embedded features, effectively improving the temporal modeling and dynamic tracing accuracy of risk propagation paths, and enhancing the reliability of risk event data relationship tracing and audit tracking.

[0015] (2) This invention achieves real-time fusion of cross-domain feature representation and message passing interaction through the network structure design of composite dual hypergraph recurrent neural network and attention fusion mechanism, which significantly improves the calculation ability of risk event propagation path reasoning and risk propagation confidence, and shows better adaptability and effect in drug clinical trial risk tracing and risk management.

[0016] (3) In terms of recursive causal link analysis of drug clinical trial risks, this invention effectively solves the problem of lack of clear end-to-end data relationship and time-directed expression in the risk event tracing process in the prior art through the recursive causal link back-reasoning algorithm. It breaks through the bottleneck of insufficient temporal and topological consistency constraints in the recursive back-reasoning of risk events in the prior art, realizes a significant improvement in the risk tracing method, and effectively enhances the computable causal chain tracing capability in the field of drug clinical trial risk management. Attached Figure Description

[0017] 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 1 This is an overall flowchart of a knowledge graph-based drug clinical trial risk traceability management method proposed in this invention; Figure 2 This is a schematic diagram of the composite dual-hypergraph recurrent neural network structure of a drug clinical trial risk tracing management method based on knowledge graph proposed in this invention; Figure 3 This is a flowchart of the recursive causal link back-reasoning algorithm for a drug clinical trial risk tracing management method based on knowledge graph proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1-3 A knowledge graph-based method for risk tracing and management in drug clinical trials, comprising: Based on multi-source heterogeneous data from drug clinical trials, an entity domain hypergraph is constructed, with trial subjects and trial elements as nodes and the interaction relationships between trial subjects and trial elements as hyperedges. Based on multi-source heterogeneous data, risk events that occur during the experiment are identified, and a relational domain hypergraph is constructed with risk events as nodes and risk propagation relationships between risk events as hyperedges. The entity domain hypergraph and the relation domain hypergraph are simultaneously input into the composite dual hypergraph recurrent neural network. Through the cross-domain interaction fusion mechanism, the node embedding of the entity domain hypergraph is updated in real time according to the hyperedge embedding of the relation domain hypergraph. Based on the node embedding of the updated entity domain hypergraph, the hyperedge embedding of the corrected relation domain hypergraph is adjusted in real time to reflect the changes in the propagation path of risk events. The cross-domain interaction and fusion mechanism is executed in a loop, so that the node embedding of the entity domain hypergraph and the hyperedge embedding of the relation domain hypergraph feed back to each other in real time in each iteration until the embedding information of the two domain hypergraphs reaches stability. Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, a risk evolution and propagation path is generated. A recursive causal link back-inference algorithm is used to extract end-to-end data relationships of risk events from the risk evolution and propagation path, and to conduct risk tracing and audit tracking in drug clinical trials.

[0020] In this embodiment, the construction of the entity domain hypergraph is specifically as follows: The trial subjects and elements are extracted from multi-source heterogeneous data of drug clinical trials. The trial subjects include subjects, research centers, and clinical trial institutions. The trial elements include drug formulations, trial visit points, and trial operation procedures. The interaction relationship between each experimental subject and its corresponding experimental element is mapped as a hyperedge, and the set of nodes connected by the multiple experimental subjects and experimental elements associated with the interaction relationship is used as the hyperedge. The aforementioned interactive relationships include: data acquisition relationships, information transmission relationships, trial visit relationships, trial operation execution relationships, drug formulation administration relationships, trial anomaly feedback relationships, and quality indicator monitoring relationships between the trial subject and trial elements; Weighted superedges are obtained by assigning weights to the superedges based on the frequency, duration, and intensity of the interaction. The weighting of the hyperedge is specifically as follows: using the frequency, duration, and intensity of interaction as evaluation indicators, constructing a pairwise comparison judgment matrix for each indicator, normalizing the judgment matrix using the analytic hierarchy process, calculating the weight corresponding to each indicator, and weighting and summing the frequency, duration, and intensity of interaction according to the weight of each indicator to obtain the comprehensive weight of the hyperedge, and assigning the comprehensive weight to the corresponding hyperedge as the weight value; The experimental subjects and elements are represented as node embeddings in a unified semantic space according to their respective experimental stages, timestamps, and associated weighted hyperedges, resulting in an entity domain hypergraph.

[0021] In this embodiment, the construction of the relational domain hypergraph is specifically as follows: Based on multi-source heterogeneous data from drug clinical trials, risk events that occur during the trial are identified, including abnormal data events, process deviation events, and events where key quality indicators exceed limits. The risk events are determined in the following manner: Abnormal data events: Statistical abnormal data detection algorithms are used to analyze the numerical distribution of data at each trial node in multi-source heterogeneous data of drug clinical trials to identify data that significantly deviate from the normal range; Process deviation events: By using rule-based process compliance analysis methods, the established trial operation procedures for drug clinical trials are compared with the actual execution process step by step to identify the operation steps that violate the trial operation procedures. Key quality indicator exceedance events: Based on the quality indicator monitoring thresholds for drug clinical trials, monitor and identify indicator values ​​that exceed the set quality standards in real time; Based on the order of occurrence and correlation of risk events, determine the propagation path between risk events, and associate risk events with direct causal propagation relationships; Each risk event is mapped to a node in the relational domain hypergraph, and the propagation path consisting of multiple risk events with direct causal propagation relationships is mapped to a hyperedge, forming a set of risk event nodes associated with risk propagation paths; Based on the correlation strength, propagation speed and impact on the overall risk of the experiment among risk events, the hyperedges of the relational domain hypergraph are weighted to obtain weighted hyperedges. The weighting of the hyperedges in the relational domain hypergraph is specifically as follows: using the correlation strength, propagation speed, and impact on the overall risk of the experiment as evaluation indicators, constructing pairwise comparison judgment matrices for each evaluation indicator, calculating and normalizing the judgment matrices using the analytic hierarchy process (AHP) to obtain the weights corresponding to each evaluation indicator, and weighting and summing the indicator values ​​of the actual correlation strength, propagation speed, and impact on the overall risk of the experiment based on the indicator weights to obtain the comprehensive weight of each hyperedge, and assigning the comprehensive weight to the corresponding hyperedge as the weight value; Based on the experimental phase, timestamp, and corresponding weighted hyperedge information at the time of the risk event, it is represented as node embedding in a unified semantic space, resulting in a relational domain hypergraph.

[0022] In this embodiment, the composite dual hypergraph recurrent neural network includes an input embedding layer, an intra-domain message passing layer, a cross-domain interaction fusion layer, an attention fusion layer, a gated recurrent update layer, and a temporal constraint and convergence output layer. The input embedding layer is used to receive node and hyperedge input data from the entity domain hypergraph and the relation domain hypergraph respectively. It generates corresponding time position codes based on the test stage and timestamp, and then fuses them with the initial features of the nodes and hyperedges to generate an initial embedding with time information. The intra-domain message passing layer is used to receive the initial embedding output from the input embedding layer. It performs bidirectional message passing of intra-domain features within the entity domain hypergraph and the relation domain hypergraph, respectively. That is, local message aggregation and diffusion are performed between nodes and hyperedges in the entity domain and between nodes and hyperedges in the relation domain, respectively, to generate entity domain node representations and relation domain hyperedge representations. The cross-domain interaction fusion layer is used to receive the entity domain node representation and relation domain hyperedge representation output by the intra-domain message passing layer. It performs bidirectional interaction of cross-domain features through the cross-domain interaction fusion mechanism, that is, cross-domain message passing from relation domain hyperedge representation to entity domain node representation, and cross-domain message passing from entity domain node representation to relation domain hyperedge representation, thereby generating a comprehensive feature representation of cross-domain interaction fusion. The attention fusion layer is used to receive the comprehensive feature representation of the cross-domain interaction fusion layer. It assigns different attention weights to the intra-domain local features and cross-domain interaction features of the entity domain and relation domain respectively through the attention mechanism, and generates fused features with dynamic update capability by weighted fusion according to the size of the attention weights. The calculation logic for the attention weight is as follows: Linear transformations are used to map the local features within the entity domain and the cross-domain interaction features between the relation domain and the entity domain, respectively, to obtain the query vector, key vector and value vector; The similarity score between the query vector and the key vector is obtained by calculating the dot product similarity between each query vector and the corresponding key vector and performing scale normalization. The similarity scores are processed by a normalized exponential function to generate normalized attention weights for each query vector on all key vectors; The normalized attention weights and their corresponding value vectors are weighted and summed to obtain the weighted feature representation. The weighted feature representations of all attention heads are concatenated and linearly transformed to obtain the fused feature; The gated loop update layer is used to receive the fusion features output by the attention fusion layer. It performs gated loop updates on the node feature embeddings of the entity domain and the hyperedge feature embeddings of the relation domain respectively, so as to realize the memory, forgetting and updating of feature representations in the time dimension, forming the currently updated embedding representation. The temporal constraint and convergence output layer is used to apply temporal directional constraints and topology consistency constraints to the embedded representations output by the gated cyclic update layer. It evaluates the magnitude of feature embedding changes after two consecutive cyclic updates. When the magnitude of the embedding changes reaches within a preset convergence threshold, the network is determined to be converged, and stable embedded representations of entity domain nodes and relation domain hyperedges are output. The time-directed constraint is as follows: Based on the actual time sequence of risk events during drug clinical trials, the message transmission direction between any two nodes in the risk propagation path generated by the network must be strictly consistent with the actual time sequence of the risk events represented by the corresponding nodes, so as to ensure that the risk propagation path strictly reflects the time causal order of events during the embedding and updating process. The topology consistency constraint is as follows: Based on the initial topology of the entity domain hypergraph and the relation domain hypergraph, it is required that the connection relationship between nodes and hyperedges during the network embedding and update process must not exceed the connection structure defined in the initial topology, and connection relationships that contradict or are inconsistent with the initial topology are not allowed, so as to ensure that the topology during the network update process is always consistent with the initial topology definition of the drug clinical trial data. By applying the two constraints mentioned above simultaneously, it is ensured that the embedding representation of entity domain nodes and relation domain hyperedges in the final network output strictly reflects the real-time causality of the risk propagation path in drug clinical trials and is consistent with the initial topology.

[0023] In this embodiment, the cross-domain interaction and fusion mechanism is specifically as follows: Based on the entity domain node representation and relation domain hyperedge representation output by the intra-domain message passing layer, cross-domain message paths from relation domain hyperedges to entity domain nodes and from entity domain nodes to relation domain hyperedges are constructed respectively. In the cross-domain message path from relational hyperedge to entity node, the relational hyperedge representation is mapped to a query vector, and the entity node representation is mapped to a key vector and a value vector. The attention weight of the relational hyperedge representation to the entity node representation is calculated through a multi-head self-attention mechanism, and the entity node representation is weighted and fused with the attention weight to form the cross-domain fusion feature of the entity node. In the cross-domain message path from entity domain nodes to relation domain hyperedges, entity domain node representations are mapped to query vectors, and relation domain hyperedge representations are mapped to key vectors and value vectors. The attention weights of entity domain node representations to relation domain hyperedge representations are calculated through a multi-head self-attention mechanism, and the relation domain hyperedge representations are weighted and fused using the attention weights to form the cross-domain fusion features of relation domain hyperedges. The cross-domain fusion features of entity domain nodes and the cross-domain fusion features of relation domain hyperedges obtained through bidirectional cross-domain message channels are merged into a comprehensive feature representation of cross-domain interaction.

[0024] In this embodiment, the hyperedge embedding of the corrected relation domain hypergraph specifically refers to: Based on the node embeddings of the updated entity domain hypergraph, calculate the change in node embeddings of the risk event node set corresponding to each relation domain hyperedge; The method for calculating the change is as follows: The difference between the node embeddings of the updated entity domain hypergraph and the node embeddings of the historical entity domain hypergraph is calculated node by node to obtain the change vector of each node embedding. Based on the set of risk event nodes corresponding to the hyperedge of the relation domain, extract the change vectors embedded in all nodes within the set of risk event nodes; The vector norm calculation method is used to calculate the magnitude of the node embedding change vector for each risk event node set; By averaging the magnitudes of the embedding change vectors of all nodes in the risk event node set, a value representing the overall embedding change of the risk event node set corresponding to each relation domain hyperedge is obtained, which is used as the change in node embedding of the risk event node set corresponding to the relation domain hyperedge. Based on the change in node embedding, calculate the adjustment weight of the current hyperedge embedding for each relation domain hyperedge; The method for calculating the adjustment weight is as follows: The change in node embedding of the risk event node set corresponding to each relation domain hyperedge is used as the computational input; The changes in the node embedding are normalized and mapped to a preset interval to obtain the normalized changes. Based on the normalized change, a threshold range is set, and the adjustment weight of the current hyperedge embedding of each relation domain hyperedge is determined using a linear function method, so that the magnitude of the adjustment weight is positively correlated with the normalized change in node embedding. The adjustment weight ranges from 0 to 1. The larger the value, the greater the adjustment required for the current relation domain hyperedge embedding. Based on the adjusted weights, the hyperedge embeddings of the relation domain hypergraph are weighted and fused to obtain the real-time corrected hyperedge embeddings of the relation domain. Based on the actual occurrence order of the risk event node set corresponding to the real-time corrected relation domain hyperedge embedding, a time-directed constraint is applied to ensure that the real-time corrected hyperedge embedding meets the temporal requirements of the risk propagation path. Perform topological consistency verification on the hyperedge embedding of the relation domain after time-directed constraint, so that the hyperedge embedding after real-time correction conforms to the initial topological structure of the hypergraph of the relation domain. The historical hyperedge embedding of the relation domain hypergraph is replaced by the hyperedge embedding of the relation domain hypergraph after the hyperedge embedding has been verified by time-directedness constraints and topological consistency, so as to complete the real-time correction of the hyperedge embedding of the relation domain hypergraph.

[0025] In this embodiment, the generation of risk evolution and propagation path specifically refers to: Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, calculate the similarity between the embedding features of each hyperedge in the relation domain hypergraph and the embedding features of the corresponding risk event node set. The similarity calculation logic is as follows: Based on the stable hypergraph and hyperedge embedding features of the relation domain and the corresponding risk event node set embedding features, the embedding features of each node in the set are averaged and fused to obtain the overall embedding features of the risk event node set. For each relation domain hyperedge embedding feature and the corresponding risk event node set overall embedding feature, calculate the cosine similarity between them. The specific method for calculating the cosine similarity is as follows: the hyperedge embedding feature of the relation domain and the overall embedding feature of the risk event node set are regarded as two vectors, the inner product of the two vectors is calculated, and then divided by the product of the magnitudes of the two vectors respectively to obtain the value of the cosine similarity between them. The cosine similarity value is used as a similarity metric for the subsequent screening and ranking of candidate risk propagation paths. Based on the calculated similarity, all relation domain hyperedges are sorted and filtered, and hyperedges exceeding the preset similarity threshold are extracted as candidate risk propagation paths; For the candidate risk propagation paths, the candidate risk propagation paths are arranged chronologically according to the timestamps corresponding to each node in the risk event node set and the actual occurrence order of the risk events. The confidence of each candidate risk propagation path is calculated by weighting the path confidence of the hyperedge similarity in the arranged candidate risk propagation paths. The confidence level is calculated using the following logic: Based on the similarity of the hyperedges of each relation domain in each candidate risk propagation path, the cumulative similarity value of the entire path is calculated by using the cumulative similarity calculation method. The cumulative similarity value of the entire path is calculated using a geometric mean based on the number of hyperedges in the path, resulting in the geometric mean of the similarity value for each path. The geometric mean of the similarity is used as the path confidence of each candidate risk propagation path. The confidence value ranges from 0 to 1, and the larger the value, the higher the confidence of the path's risk propagation. The confidence level of each path in the candidate risk propagation path is filtered according to the preset path confidence threshold, and the paths with confidence levels higher than the preset threshold are retained to obtain the risk evolution and propagation path. The risk evolution and propagation path is represented by sequentially associating risk event nodes with their corresponding timestamps, generating a time-series risk evolution and propagation path for risk tracing and audit tracking.

[0026] In this embodiment, the recursive causal link back-reasoning algorithm is specifically as follows: Based on the risk evolution and propagation path, target risk event nodes to be traced are selected, and an initial recursive node set is established based on the embedding characteristics and corresponding timestamps of the target risk event nodes. Based on the target risk event node in the initial recursive node set, the predecessor risk event node that is directly related to the target risk event node and has an earlier timestamp is determined in the risk evolution and propagation path with the target risk event node as the endpoint, and the predecessor node set of the current recursive level is established. Based on the node embedding features of the predecessor node set at the current recursive level, calculate the cosine similarity between the node embedding features of each node in the predecessor node set and the target risk event node embedding features, and determine the nodes with a cosine similarity higher than the preset cosine similarity threshold as valid predecessor nodes. For valid precursor nodes, based on the embedded features and corresponding timestamps, we recursively backtrack to earlier moments in the risk evolution and propagation path, identifying upstream precursor risk event nodes layer by layer, and using the similarity threshold of node embedded features for layer-by-layer screening during the recursive backtracking process. During the recursive backward process, topological consistency verification is performed on each recursive path to ensure that the connection structure of the nodes in the backward path is consistent with the topological structure of the risk evolution and propagation path. Paths that do not meet the topological consistency requirements are terminated in a timely manner. During the recursive reverse path formation process, the path confidence is calculated by multiplying the embedding feature similarity of consecutive nodes in each recursive path, and a path confidence threshold is used to filter low confidence paths, retaining only reverse paths with a path confidence higher than the path confidence threshold. The nodes in the retained reverse path are arranged in chronological order according to their respective timestamps to obtain a recursive causal link with time order, forming an end-to-end data relationship from the target risk event node to the initial risk event node. The above recursive backward process is repeated for each target risk event node until the end-to-end data relationship link of all target risk event nodes in drug clinical trial risk tracing is completed, and the recursive causal link of all risk events is output for drug clinical trial risk tracing and audit tracking. Example

[0027] To verify the feasibility of this invention in practice, it was applied to a drug clinical trial risk management scenario, specifically a Phase III clinical trial of an investigational drug. The trial lasted 12 months, involved 500 subjects and 10 research centers, generating approximately 400,000 pieces of multi-source heterogeneous clinical data, including subject visit records, research center monitoring logs, drug formulation operation records, abnormal data and process deviation event records, and key quality indicator monitoring data. Traditional methods are generally based on manual verification and statistical analysis. The identification and tracing of risk events mainly rely on human experience, making it difficult to achieve automated tracing and to construct a clear causal propagation path for risk events. This results in low risk management efficiency, poor tracing accuracy, and the inability to conduct real-time tracking and auditing.

[0028] During implementation, the system first collects all multi-source heterogeneous data from drug clinical trials. Based on the data's content, the system automatically extracts the trial subjects, including subjects, research centers, clinical trial institutions, and trial elements, including drug formulations, trial visit nodes, and trial operation procedures. Subsequently, the interactive relationships between trial subjects and trial elements, such as data collection relationships, visit relationships, and abnormal feedback relationships, are mapped as hyperedges, forming the basis for constructing an entity domain hypergraph. Simultaneously, based on risk events recorded during the trial, such as abnormal data events, process deviation events, and key quality indicator exceedance events, a relationship domain hypergraph is constructed, using risk events as nodes and mapping the propagation paths between risk events as hyperedges.

[0029] The two hypergraphs mentioned above are input into a composite dual-hypergraph recurrent neural network. Through a cross-domain interaction fusion mechanism, embedded feature information is fed back between the entity domain and the relation domain in real time. The input embedding layer first generates initial features based on the experimental stage and timestamp. The intra-domain message passing layer performs message passing on the entity domain and relation domain respectively and generates corresponding feature representations. The cross-domain interaction fusion layer then performs bidirectional cross-domain message interaction and generates comprehensive features. The attention fusion layer calculates the attention weights of local and cross-domain features through an attention mechanism, and outputs the fused feature based on these weights to the gated recurrent update layer to achieve real-time updates of entity domain node and relation domain hyperedge embeddings. Subsequently, the temporal constraint and convergence output layer determines the stability of the embedded information based on the feature change magnitude. When the feature change magnitude is below 0.001 twice consecutively, the network is considered converged and outputs stable node and hyperedge embedding information.

[0030] After obtaining stable node and hyperedge embeddings, the system generates risk evolution and propagation paths based on embedding feature similarity, and filters for valid paths using a path confidence threshold of 0.75. A recursive causal link back-inference algorithm is further employed, starting with the target risk event node, and recursively filtering for valid predecessor nodes layer by layer based on the node embedding feature cosine similarity threshold of 0.85. Verification is performed using topological consistency and temporal consistency constraints, ultimately generating an end-to-end causal link with a clear temporal sequence. Table 1 shows a comparison of predicted paths and actual source tracing paths for some risk event nodes in this embodiment.

[0031] Table 1 Comparison of Predicted Risk Event Pathways and Actual Tracing Pathways in Drug Clinical Trials

[0032] As can be observed from the data in Table 1, the method of this invention highly matches the actual risk propagation path prediction, with a node coverage rate exceeding 85% and a confidence level above 0.85 for the vast majority of paths. Specifically, for risk event FX-001, the actual tracing path had 8 nodes, and the predicted path also accurately matched 8 nodes, achieving a node coverage rate of 100%, a path confidence level of 0.92, and an average cosine similarity of 0.94, demonstrating the high accuracy of this method for tracing the source of risk events. For event FX-025, there was a one-node difference between the predicted and actual number of nodes, but the node coverage rate was still as high as 85.7%, and the path confidence level also reached 0.85, indicating that even with local data noise, this invention can still effectively ensure the overall accuracy of the tracing path.

[0033] Through detailed verification in this embodiment, the method of the present invention effectively solves the problems of low efficiency and insufficient accuracy of risk tracing in the existing technology under multi-source heterogeneous clinical data environment. It realizes clear causal chain calculation of risk events, real-time tracing and accurate audit tracking, significantly improves the efficiency and reliability of drug clinical trial risk management, and has good practical application value.

[0034] The above description is only a preferred embodiment 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 knowledge graph-based method for risk traceability management in drug clinical trials, characterized in that, include: Based on multi-source heterogeneous data from drug clinical trials, an entity domain hypergraph is constructed, with trial subjects and trial elements as nodes and the interaction relationships between trial subjects and trial elements as hyperedges. Based on multi-source heterogeneous data, risk events that occur during the experiment are identified, and a relational domain hypergraph is constructed with risk events as nodes and risk propagation relationships between risk events as hyperedges. The entity domain hypergraph and the relation domain hypergraph are simultaneously input into the composite dual hypergraph recurrent neural network. Through the cross-domain interaction fusion mechanism, the node embedding of the entity domain hypergraph is updated in real time according to the hyperedge embedding of the relation domain hypergraph. Based on the node embedding of the updated entity domain hypergraph, the hyperedge embedding of the corrected relation domain hypergraph is performed in real time. The cross-domain interaction and fusion mechanism is executed in a loop, so that the node embedding of the entity domain hypergraph and the hyperedge embedding of the relation domain hypergraph feed back to each other in real time in each iteration until the embedding information of the two domain hypergraphs reaches stability. Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, a risk evolution and propagation path is generated. A recursive causal link back-inference algorithm is used to extract end-to-end data relationships of risk events from the risk evolution and propagation path, and to conduct risk tracing and audit tracking in drug clinical trials.

2. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The construction of the entity domain hypergraph is specifically as follows: The trial subjects and elements are extracted from multi-source heterogeneous data of drug clinical trials. The trial subjects include subjects, research centers, and clinical trial institutions. The trial elements include drug formulations, trial visit points, and trial operation procedures. The interaction relationship between each experimental subject and its corresponding experimental element is mapped as a hyperedge, and the set of nodes connected by the multiple experimental subjects and experimental elements associated with the interaction relationship is used as the hyperedge. Weighted superedges are obtained by assigning weights to the superedges based on the frequency, duration, and intensity of the interaction. The experimental subjects and elements are represented as node embeddings in a unified semantic space according to their respective experimental stages, timestamps, and associated weighted hyperedges, resulting in an entity domain hypergraph.

3. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The construction of the relational domain hypergraph is specifically as follows: Based on multi-source heterogeneous data from drug clinical trials, risk events that occur during the trial are identified, including abnormal data events, process deviation events, and events where key quality indicators exceed limits. Based on the order of occurrence and correlation of risk events, determine the propagation path between risk events, and associate risk events with direct causal propagation relationships; Each risk event is mapped to a node in the relational domain hypergraph, and the propagation path consisting of multiple risk events with direct causal propagation relationships is mapped to a hyperedge, forming a set of risk event nodes; Based on the correlation strength, propagation speed and impact on the overall risk of the experiment among risk events, the hyperedges of the relational domain hypergraph are weighted to obtain weighted hyperedges. Based on the experimental phase, timestamp, and corresponding weighted hyperedge information at the time of the risk event, it is represented as node embedding in a unified semantic space, resulting in a relational domain hypergraph.

4. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The composite dual hypergraph recurrent neural network includes an input embedding layer, an intra-domain message passing layer, a cross-domain interaction fusion layer, an attention fusion layer, a gated recurrent update layer, and a temporal constraint and convergence output layer. The input embedding layer is used to receive node and hyperedge input data from the entity domain hypergraph and the relation domain hypergraph respectively. It generates corresponding time position codes based on the test stage and timestamp, and then fuses them with the initial features of the nodes and hyperedges to generate an initial embedding with time information. The intra-domain message passing layer is used to receive the initial embedding output from the input embedding layer. It performs bidirectional message passing of intra-domain features within the entity domain hypergraph and the relation domain hypergraph, respectively, and generates entity domain node representations and relation domain hyperedge representations. The cross-domain interaction fusion layer is used to receive entity domain node representations and relation domain hyperedge representations output by the intra-domain message passing layer. Through the cross-domain interaction fusion mechanism, it performs bidirectional interaction of cross-domain features to generate comprehensive feature representations. The attention fusion layer is used to receive the comprehensive feature representation of the cross-domain interaction fusion layer. It assigns different attention weights to the intra-domain local features and cross-domain interaction features of the entity domain and the relation domain respectively through the attention mechanism, and generates fused features by weighting and fusing according to the size of the attention weights. The gated loop update layer is used to receive the fusion features output by the attention fusion layer, and performs gated loop updates on the node feature embeddings of the entity domain and the hyperedge feature embeddings of the relation domain respectively, forming the updated embedding representation at the current time. The temporal constraint and convergence output layer is used to apply temporal directional constraints and topological consistency constraints to the embedded representations output by the gated cyclic update layer. It evaluates the magnitude of feature embedding changes after two consecutive cyclic updates. When the magnitude of the embedding changes reaches within a preset convergence threshold, the network is determined to be converged, and stable embedding representations of entity domain nodes and relation domain hyperedges are output.

5. The method for risk tracing and management of drug clinical trials based on knowledge graphs according to claim 4, characterized in that, The cross-domain interaction and fusion mechanism is specifically as follows: Based on the entity domain node representation and relation domain hyperedge representation output by the intra-domain message passing layer, cross-domain message paths from relation domain hyperedges to entity domain nodes and from entity domain nodes to relation domain hyperedges are constructed respectively. In the cross-domain message path from relational hyperedge to entity node, the relational hyperedge representation is mapped to a query vector, and the entity node representation is mapped to a key vector and a value vector. The attention weight of the relational hyperedge representation to the entity node representation is calculated through a multi-head self-attention mechanism, and the entity node representation is weighted and fused with the attention weight to form the cross-domain fusion feature of the entity node. In the cross-domain message path from entity domain nodes to relation domain hyperedges, entity domain node representations are mapped to query vectors, and relation domain hyperedge representations are mapped to key vectors and value vectors. The attention weights of entity domain node representations to relation domain hyperedge representations are calculated through a multi-head self-attention mechanism, and the relation domain hyperedge representations are weighted and fused using the attention weights to form the cross-domain fusion features of relation domain hyperedges. The cross-domain fusion features of entity domain nodes and the cross-domain fusion features of relation domain hyperedges obtained through bidirectional cross-domain message channels are merged into a comprehensive feature representation of cross-domain interaction.

6. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The hyperedge embedding of the corrected relation domain hypergraph is specifically as follows: Based on the node embeddings of the updated entity domain hypergraph, calculate the change in node embeddings of the risk event node set corresponding to each relation domain hyperedge; Based on the change in node embedding, calculate the adjustment weight of the current hyperedge embedding for each relation domain hyperedge; Based on the adjusted weights, the hyperedge embeddings of the relation domain hypergraph are weighted and fused to obtain the real-time corrected hyperedge embeddings of the relation domain. Based on the actual occurrence order of the risk event node set corresponding to the hyperedge embedding of the relation domain after real-time correction, a time-directed constraint is applied; Perform topological consistency verification on hyperedge embeddings of relation domains subject to time-directed constraints; The historical hyperedge embedding of the relation domain hypergraph is replaced by the hyperedge embedding of the relation domain hypergraph after the hyperedge embedding has been verified by time-directedness constraints and topological consistency, so as to complete the real-time correction of the hyperedge embedding of the relation domain hypergraph.

7. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The generated risk evolution and propagation path is specifically as follows: Based on the node embeddings of the stabilized entity domain hypergraph and the hyperedge embeddings of the relation domain hypergraph, calculate the similarity between the embedding features of each hyperedge in the relation domain hypergraph and the embedding features of the corresponding risk event node set. Based on the calculated similarity, all relation domain hyperedges are sorted and filtered, and hyperedges exceeding the preset similarity threshold are extracted as candidate risk propagation paths; For the candidate risk propagation paths, the candidate risk propagation paths are arranged chronologically according to the timestamps corresponding to each node in the risk event node set and the actual occurrence order of the risk events. The confidence of each candidate risk propagation path is calculated by weighting the hyperedge similarity of the arranged candidate risk propagation paths with path confidence. The confidence level of each path in the candidate risk propagation path is filtered according to the preset path confidence threshold, and the paths with confidence levels higher than the preset threshold are retained to obtain the risk evolution and propagation path. The risk evolution and propagation path is represented by sequentially associating risk event nodes with their corresponding timestamps, thus generating a time-series risk evolution and propagation path.

8. The method for risk traceability management of drug clinical trials based on knowledge graphs according to claim 1, characterized in that, The recursive causal link back-reasoning algorithm is as follows: Based on the risk evolution and propagation path, select the target risk event nodes to be traced and establish an initial recursive node set; Based on the target risk event node in the initial recursive node set, determine the predecessor risk event node that is directly related to the target risk event node and has an earlier timestamp, and establish the predecessor node set for the current recursive level. Based on the node embedding features of the predecessor node set at the current recursive level, calculate the cosine similarity between the node embedding features of each node in the predecessor node set and the target risk event node embedding features, and determine the nodes with a cosine similarity higher than the preset cosine similarity threshold as valid predecessor nodes. For valid precursor nodes, based on the embedded features and corresponding timestamps, we recursively backtrack to earlier moments in the risk evolution and propagation path, identifying upstream precursor risk event nodes layer by layer, and using the similarity threshold of node embedded features for layer-by-layer screening during the recursive backtracking process. During the recursive reverse process, topological consistency is verified for each recursive path, and recursion is terminated in a timely manner for paths that do not meet the topological consistency requirements. During the recursive backward path formation process, the path confidence is calculated by multiplying the embedding feature similarity of consecutive nodes in each recursive path, and only backward paths with a path confidence threshold are retained. The nodes in the retained reverse path are arranged in chronological order according to their respective timestamps to obtain a recursive causal link with chronological order. The above recursive backward process is repeated for each target risk event node until the end-to-end data relationship link of all target risk event nodes in drug clinical trial risk tracing is completed, and the recursive causal link of all risk events is output for drug clinical trial risk tracing and audit tracking.

Images