A method and system for predicting the association of adverse events related to concomitant medication in clinical trial monitoring

Abstract

This invention discloses a method and system for predicting adverse events associated with concomitant medications in clinical trial monitoring, relating to the fields of artificial intelligence and biomedical information processing. The method acquires and standardizes patient data, concomitant medication data, adverse event data, disease data, and drug target data to construct a heterogeneous graph of concomitant medication adverse events, including patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes. Different types of nodes are embedded, with patient nodes and their concomitant medication combinations as initial constraints and adverse event nodes as target constraints, and target meta-paths are determined through adaptive meta-path learning. Then, a risk representation under the concomitant medication combination is generated based on a hierarchical heterogeneous graph attention network, predicting adverse events and generating risk scores, while simultaneously outputting the results of risk causation tracing and evidence chain presentation. This invention can collaboratively represent multiple types of associations, improving the accuracy and interpretability of risk prediction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and biomedical information processing technology, specifically to a method and system for predicting adverse events associated with concomitant medications in clinical trial monitoring. In particular, it relates to a method and system for predicting adverse events associated with concomitant medications based on heterogeneous graph construction, adaptive meta-path learning, and hierarchical heterogeneous graph attention networks, which collaboratively models patient data, concomitant medication data, adverse event data, disease data, and drug target data in clinical trials, and generates adverse event risk scores and interpretable prediction results. Background Technology

[0002] In drug clinical trials, subjects typically have specific disease states, prior treatment histories, and individualized medication backgrounds. In addition to the investigational drug, subjects may also use multiple concomitant medications, such as medications for underlying diseases, supportive care, or prophylactic medications. Complex relationships may form between concomitant medications and the investigational drug, between different concomitant medications, and between the drug and the patient's disease state or drug target, thus affecting the risk of adverse events. Therefore, clinical trial monitoring requires comprehensive analysis of multiple types of information, including patient information, concomitant medications, adverse events, disease information, and drug target information, to assess the risk of adverse events for a specific subject with a particular combination of concomitant medications.

[0003] Existing technologies have explored methods for predicting adverse drug reactions or side effects of drug combinations using machine learning or graph neural networks. One type of method constructs a heterogeneous graph containing nodes such as patients, diseases, drugs, and adverse drug reactions to predict patient-level adverse drug reactions. Another type constructs a multimodal graph relating drugs, protein targets, and drug combination side effects to predict potential side effects from drug combinations. Still other methods predict the existence of associations between combined drugs and side effect nodes based on meta-pathways or heterogeneous information networks. These methods can, to some extent, utilize graph structures to express the relationships between medical entities, thereby improving the ability to predict drug safety risks.

[0004] However, existing technologies still have shortcomings in the context of monitoring concomitant medications in clinical trials. First, patient-level adverse drug reaction prediction methods typically focus on modeling the overall association between patients, diseases, drugs, and adverse reactions, failing to fully integrate the concomitant medication combination itself as a risk initiation object, and failing to form a targeted association path starting from the patient and their concomitant medication combination around specific adverse event nodes. Second, drug combination side effect prediction methods typically use drug pairs or drug combinations as the main prediction objects, focusing on determining whether there is a correlation between the drug combination and side effects, making it difficult to simultaneously combine patient disease background, drug target relationships, and individual information of clinical trial participants for patient-level risk assessment. Third, while prediction methods based on knowledge graphs or meta-paths can express various medical entity relationships, their meta-paths are usually used to enhance node representations or edge predictions, lacking a mechanism to generate candidate meta-paths and screen target meta-paths by using patient nodes and their concomitant medication combinations as initial constraints and adverse event nodes as target constraints.

[0005] Furthermore, the risk of adverse events in clinical trials is often not determined independently by a single drug, disease, or target, but rather by the combined effects of multiple related factors. For example, a patient's underlying disease or comorbidities may alter their susceptibility to certain types of adverse events; concomitant medication combinations may alter drug exposure or efficacy through drug interactions; the relationship between drug targets and disease mechanisms may influence the tendency for adverse events to occur; and existing drug-adverse event associations may provide direct evidence for risk prediction. These relationships exhibit chain-like connections and cross-influences. If modeling is only performed using patient-drug relationships, drug-adverse event relationships, drug-target relationships, or drug interaction relationships separately, without generating and screening target meta-paths connecting patients, concomitant medication combinations, and adverse events within a unified heterogeneous graph, it becomes difficult to accurately identify key drug nodes, key drug combinations, and key association paths in the risk formation process.

[0006] Furthermore, clinical trial monitoring places high demands on the interpretability of prediction results. While some existing deep learning prediction methods can output the probability of adverse events or risk scores, there is a lack of clear correspondence between the prediction results and specific nodes, paths, and their contribution relationships in the graph. For clinical trial monitors, if they cannot clearly identify which drug nodes, drug combinations, and target meta-paths the risk score primarily originates from, and how different target meta-paths are integrated to form the final risk representation, it is difficult to use the prediction results for risk causation tracing, medical verification, and monitoring record retention.

[0007] Therefore, existing technologies still require a method for predicting adverse events associated with concomitant medications in clinical trial monitoring. This method should be able to generate candidate meta-paths and determine target meta-paths in a heterogeneous graph containing patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, using patient nodes and their concomitant medication combinations as initial constraints and adverse event nodes as target constraints. Furthermore, it should determine the importance of associated nodes, the importance of target meta-paths, and the fusion weights of different target meta-paths through node-level attention, meta-path-level attention, and global attention, respectively. This would allow the method to predict adverse event risks while simultaneously outputting results on risk causal tracing and evidence chain presentation, thereby improving the accuracy and interpretability of adverse event risk prediction for concomitant medications in clinical trials. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for predicting concomitant medication adverse events in clinical trial monitoring. This addresses the problem in existing technologies where patient-level adverse reaction prediction, drug combination side effect prediction, drug target relationship modeling, and clinical trial concomitant medication association judgment are fragmented and lack synergistic characterization of multiple types of associations between patients, concomitant medication combinations, adverse events, diseases, and drug targets. By constructing a heterogeneous graph of concomitant medication adverse events and determining key nodes, key drug combinations, and target meta-paths based on adaptive meta-path learning and hierarchical heterogeneous graph attention networks, the accuracy and interpretability of concomitant medication adverse event risk prediction in clinical trials are improved.

[0009] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: A method for predicting concomitant adverse events related to medication use in clinical trial monitoring, comprising: Acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data from clinical trials, and standardize drug names, adverse event terminology, and disease information; A heterogeneous graph of adverse drug events is constructed based on standardized data. The heterogeneous graph includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations. Node embedding is performed on different types of nodes in the heterogeneous graph to obtain node representations in a unified vector space; Using patient nodes and their accompanying medication combinations as initial constraints and adverse event nodes as target constraints, candidate meta-paths are generated from the heterogeneous graph based on adaptive meta-path learning, and target meta-paths are determined. The target meta-paths represent the association between the accompanying medication combinations and adverse events through disease nodes, target nodes, or drug interactions. Based on a hierarchical heterogeneous graph attention network, feature learning is performed on the target meta-path and its associated nodes to obtain a risk representation of the patient under the combination of concomitant medications. Based on the risk representation, predict the adverse events that the accompanying medication combination may induce and generate a risk score; The hierarchical heterogeneous graph attention network determines the importance of associated nodes, the importance of target meta-paths, and the fusion weights of different target meta-paths based on node-level attention, meta-path-level attention, and global attention, respectively. Based on the importance of associated nodes, the importance of target meta-paths, and the fusion weights, it determines key drug nodes, key drug combinations, and corresponding target meta-paths, and outputs interpretable prediction results including risk causal tracing and evidence chain display.

[0010] Optionally, the standardization process includes: standardizing the drug names in the accompanying medication data into drug codes, standardizing the adverse event terms in the adverse event data into standard adverse event terms, standardizing the disease names in the disease data into disease codes, and mapping the patient data, drug codes, standard adverse event terms, disease codes, and drug target data to corresponding patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, respectively.

[0011] Optionally, the patient node is used to characterize individual patient characteristics, the drug node is used to characterize the investigational drug and concomitant drugs, the adverse event node is used to characterize historically recorded adverse events or the concept of adverse events to be predicted, the disease node is used to characterize underlying diseases or comorbidities, and the target node is used to characterize the drug's target. The drug-drug relationship connects the patient node and the drug node, the disease relationship connects the patient node and the disease node, the drug-drug interaction relationship connects different drug nodes, the drug-target relationship connects the drug node and the target node, the target-disease relationship connects the target node and the disease node, and the drug-adverse event association relationship connects the drug node and the adverse event node.

[0012] Optionally, the node embedding includes: generating patient node representations based on patient clinical characteristics, generating drug node representations based on drug chemical structure features and pharmacological features, generating adverse event node representations based on adverse event terminology semantics, generating disease node representations based on disease coding hierarchy, and generating target node representations based on target sequence features or functional features, so that the node representations of different types of nodes are in the same dimension vector space.

[0013] Optionally, the adaptive meta-path learning includes: using the current patient node and at least two drug nodes connected to it as the starting structure of the path, using the concept node of the adverse event to be predicted in the global heterogeneous graph as the target structure of the path, generating candidate meta-paths based on the drug relationship, disease relationship, drug interaction relationship, drug-target relationship, target-disease relationship and drug-adverse event association relationship in the heterogeneous graph, and determining the target meta-path according to the contribution of each candidate meta-path to the prediction of adverse events.

[0014] Optionally, the adaptive meta-path learning is performed by a path search agent; during the training phase, the state of the path search agent includes the current node representation, the target adverse event node representation, and the generated path representation; the actions of the path search agent include selecting the next neighbor node or the next neighbor node type from the current node; the reward of the path search agent is determined based on the downstream adverse event prediction performance improvement and the path length penalty term; during the inference phase, the path search agent determines the target meta-path based on the path search strategy obtained through training.

[0015] Optionally, the target meta-path includes at least two of the following: a direct medication path where the patient node is connected to the adverse event node via a drug node; a drug interaction path where the patient node is connected to the adverse event node via at least two drug nodes with drug interaction relationships; and a disease-target association path where the patient node is connected to the adverse event node via a disease node, a target node, and a drug node.

[0016] Optionally, the node-level attention is used to determine the node contribution of drug nodes, disease nodes, target nodes, and adverse event nodes in the same target meta-path to adverse event prediction; the meta-path-level attention is used to determine the path contribution of different target meta-paths to adverse event prediction; and the global attention is used to combine the patient node representation, the concomitant medication combination representation, and the path contribution to determine the fusion weight of different target meta-paths in the current patient prediction task, so as to generate the risk representation of the patient under the concomitant medication combination.

[0017] Optionally, the risk score is generated based on the predicted probability of adverse events; the interpretable prediction results include predicted adverse events, risk scores, key drug nodes, key drug combinations, target metapaths, importance of associated nodes, importance of target metapaths, and evidence chain presentation results. The key drug nodes are determined based on node contribution, the key drug combinations are determined based on the contribution of drug nodes and path in the drug interaction path, and the evidence chain presentation results display at least some of the nodes and their corresponding contributions from patient nodes, drug nodes, disease nodes, target nodes, and adverse event nodes in the node order of the target metapath.

[0018] To achieve the above-mentioned technical objectives, the present invention also adopts the following technical solution: A predictive system for concomitant adverse events related to medication use in clinical trial monitoring, comprising: The data standardization module is used to acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data in clinical trials, and to standardize drug names, adverse event terms, and disease information. The heterogeneous graph construction module is used to construct a heterogeneous graph of accompanying adverse drug events based on standardized data. The heterogeneous graph includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-drug interaction relationships, drug-target relationships, target-disease relationships, and drug-adverse event association relationships. The node embedding module is used to embed different types of nodes in the heterogeneous graph to obtain node representations in a unified vector space. The meta-path learning module is used to generate candidate meta-paths and determine target meta-paths from the heterogeneous graph, using the current patient node and at least two drug nodes connected to it as the path starting structure and the adverse event concept node to be predicted in the global heterogeneous graph as the path target structure. The hierarchical heterogeneous graph attention module is used to learn features of the target meta-path and its associated nodes based on node-level attention, meta-path-level attention and global attention, so as to obtain the risk representation of the patient under the combination of concomitant medications. The risk prediction module is used to obtain the predicted probability of adverse events based on the risk representation, and to generate a risk score result based on the predicted probability of adverse events. The interpretable output module is used to determine key drug nodes, key drug combinations and corresponding target meta-paths based on the importance of associated nodes, the importance of target meta-paths and the fusion weight, and output interpretable prediction results including risk cause tracing and evidence chain display.

[0019] The main advantages of this invention compared to existing technologies are as follows: This invention constructs patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes in the same heterogeneous graph of concomitant medication adverse events. It generates candidate meta-paths and target meta-paths with patient nodes and their concomitant medication combinations as starting constraints and adverse event nodes as target constraints. This can collaboratively represent the multi-type associations between individual patient characteristics, concomitant medication combinations, disease background, drug targets, and adverse events, avoiding the problem of existing patient-level adverse reaction prediction, drug combination side effect prediction, drug target modeling, and combined medication association judgment being mutually isolated.

[0020] This invention determines the target meta-path from the heterogeneous graph through adaptive meta-path learning, so that adverse event risk prediction no longer relies solely on full graph aggregation or fixed relationship modeling, but can perform feature learning around the actual association path between the combination of concomitant medications and adverse events, which is beneficial to improving the pertinence of risk association identification in complex concomitant medication scenarios.

[0021] This invention determines the importance of associated nodes, the importance of target meta-paths, and the fusion weights of different target meta-paths through node-level attention, meta-path-level attention, and global attention, respectively. It can establish a correspondence between node contribution, path contribution, and overall risk representation, thereby improving the accuracy of adverse event risk representation for patients with concomitant medication combinations.

[0022] This invention predicts adverse events that may be induced by concomitant drug combinations based on risk representation and generates risk scores. At the same time, it determines key drug nodes, key drug combinations and corresponding target meta-paths based on the importance of associated nodes, the importance of target meta-paths and fusion weights. This enables the risk prediction results corresponding to the risk scores to have a clear graph structure interpretation source, reducing the problem that simple probability output is difficult to interpret.

[0023] This invention outputs interpretable predictive results that include risk cause tracing and evidence chain presentation, enabling clinical trial monitors to know the key drugs, key drug combinations, target metapaths and their contribution relationships corresponding to the risk score, which is beneficial for subsequent risk verification, medical judgment and monitoring record retention. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the method steps of the method for predicting the association of adverse events with accompanying medication according to the present invention. Figure 2 This is a system architecture diagram of the adverse drug event association prediction system of the present invention; Figure 3 This is a heterogeneous diagram of accompanying adverse drug events and a schematic diagram of the target element path of the present invention; Figure 4 This is a schematic diagram of the hierarchical heterogeneous graph attention and interpretable output of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0026] This invention provides a method and system for predicting adverse events associated with concomitant medications in clinical trial monitoring. The method and system process patient data, concomitant medication data, adverse event data, disease data, and drug target data from clinical trials. By constructing a heterogeneous graph of concomitant medication adverse events and based on adaptive meta-path learning and a hierarchical heterogeneous graph attention network, it determines the risk representation of adverse events for patients under concomitant medication combinations, and then outputs the predicted probability of adverse events, risk score, key drug nodes, key drug combinations, target meta-paths, risk trigger source tracing, and evidence chain presentation results.

[0027] In this embodiment of the invention, patient data may include patient age, gender, underlying diseases, comorbidities, previous treatment history, clinical trial group, laboratory test results, or other individual patient characteristics. Concomitant medication data may include the investigational drug, concomitant medications, supportive care medications, drug name, drug category, dosage, frequency of administration, route of administration, or medication records. Adverse event data may include the name of the adverse event, time of occurrence, severity, outcome, treatment measures, or related medical records. Disease data may include the investigational indication, underlying diseases, comorbidities, or disease codes. Drug target data may include the drug's target, target function, target-related disease mechanisms, or drug-target association information.

[0028] like Figure 2 As shown, the adverse event association prediction system for clinical trial monitoring provided by this embodiment of the invention includes a data standardization module, a heterogeneous graph construction module, a node embedding module, a meta-path learning module, a hierarchical heterogeneous graph attention module, a risk prediction module, and an interpretable output module.

[0029] The modules described above process data in the following order: data input, standardization, heterogeneous graph construction, node embedding, meta-path learning, hierarchical attention feature learning, risk prediction, and interpretable output. Through the collaborative work of these modules, the system can transform various relationships between patients, concomitant medication combinations, adverse events, diseases, and drug targets into a computable graph structure, and further generate interpretable risk prediction results.

[0030] The data standardization module is used to acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data in clinical trials, and to standardize drug names, adverse event terms, and disease information.

[0031] In one implementation, the data standardization module can acquire data from clinical trial data management systems, electronic data capture systems, drug databases, medical terminology databases, drug target databases, or historical clinical trial databases. For data from different sources, the data standardization module first performs field identification and entity identification, and then maps the identified drug, adverse event, and disease entities to a unified standard.

[0032] For concomitant medication data, the data standardization module can standardize the generic name, brand name, abbreviation, or different language expressions of a drug into a drug code. The drug code can be an ATC code or any other code that can represent a drug category or drug entity. The standardized drug code is used to generate or match drug nodes.

[0033] For adverse event data, the data standardization module can standardize the names of adverse events in clinical trial records into standard adverse event terminology. For example, it can map adverse event names with different expressions to unified standard terminology, so that the same type of adverse event can be associated with the same adverse event node.

[0034] For disease data, the data standardization module can standardize underlying diseases, comorbidities, trial indications, or disease diagnosis names into disease codes. These standardized disease codes are then used to generate or match disease nodes.

[0035] For patient data and drug target data, the data standardization module maps individual patient characteristics to patient nodes and drug target data to target nodes or drug-target relationships. Through this process, patient data, drug codes, adverse event standard terms, disease codes, and drug target data are associated with patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, respectively.

[0036] The heterogeneity graph construction module is used to build heterogeneity graphs of adverse drug events based on standardized data. For example... Figure 3 As shown, the heterogeneous graph of adverse drug events includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations.

[0037] Patient nodes are used to characterize participants in clinical trials. Patient nodes can carry individual patient characteristics such as age, gender, underlying diseases, comorbidities, previous treatment history, clinical trial group, and laboratory test results.

[0038] Drug nodes are used to characterize investigational drugs and concomitant medications. Drug nodes can carry drug codes, drug categories, chemical structural characteristics, pharmacological characteristics, dosing information, mechanisms of action, or information associated with known adverse events.

[0039] Adverse event nodes are used to characterize historically recorded adverse event standard concepts or adverse event standard concepts to be predicted. For historically recorded adverse event standard concepts, the adverse event node may carry the adverse event standard terminology, severity, frequency of occurrence, outcome information, or historical occurrence details; for adverse event standard concepts to be predicted, the adverse event node is used as a target node for predicting the adverse event risk of the current patient, and does not indicate that the corresponding adverse event has already occurred in the current patient.

[0040] Disease nodes are used to characterize a patient's underlying disease, comorbidities, or indications for the trial. Disease nodes can carry disease codes, disease levels, disease severity, or disease mechanism information.

[0041] Target nodes are used to characterize drug targets. Target nodes can carry information such as target name, protein sequence, target function, involved biological pathways, or the association between the target and disease.

[0042] In heterogeneous graphs, drug-use relationships connect patient nodes and drug nodes to indicate that a patient is using a particular drug; disease relationships connect patient nodes and disease nodes to indicate that a patient has a certain underlying disease, comorbidity, or indication for the trial; drug-interaction relationships connect different drug nodes to indicate that there is an interaction or potential interaction between different drugs; drug-target relationships connect drug nodes and target nodes to indicate that the drug acts on the corresponding target; target-disease relationships connect target nodes and disease nodes to indicate that there is a correlation between the target and the disease mechanism; and drug-adverse event association relationships connect drug nodes and adverse event nodes to indicate that there is a known, suspected, or historically statistical association between the drug and the adverse event.

[0043] In this embodiment, the heterogeneous graph of adverse events related to concomitant medications may include a historical known association network and a current patient subgraph. The historical known association network is used to represent the nodes and relationships already existing in historical clinical trial data, drug knowledge data, disease knowledge data, drug target data, and historical drug-adverse event association data; the current patient subgraph is used to represent the current patient node to be predicted, the concomitant medication combination node connected to the current patient node, the disease node, and the related target node. For the adverse event to be predicted, the system uses the corresponding standard concept node of the adverse event as the target node in the heterogeneous graph formed by the historical known association network and the current patient subgraph, and predicts the possibility of a potential association between the current patient and its concomitant medication combination and the target node.

[0044] In this way, the heterogeneous graph construction module integrates scattered clinical trial data, drug information, disease information, adverse event information, and drug target information into a single graph structure, enabling subsequent models to analyze the potential association between concomitant drug combinations and adverse events from multiple relational dimensions.

[0045] The node embedding module is used to embed different types of nodes in the heterogeneous graph of adverse drug events to obtain node representations in a unified vector space.

[0046] Since the data sources, field structures, and semantic meanings of patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes are different, the node embedding module extracts corresponding features for different types of nodes and then maps each type of node to a vector space of the same dimension.

[0047] For patient nodes, the node embedding module can generate patient node representations based on the patient's clinical characteristics such as age, gender, underlying diseases, comorbidities, clinical trial group, previous treatment history, and laboratory test results.

[0048] For drug nodes, the node embedding module can generate drug node representations based on drug chemical structure characteristics, pharmacological characteristics, drug category, mechanism of action, administration information, metabolic pathway, or known adverse event association information.

[0049] For adverse event nodes, the node embedding module can generate adverse event node representations based on adverse event standard terms, terminology levels, semantic similarity, or historical adverse event records.

[0050] For disease nodes, the node embedding module can generate disease node representations based on disease codes, disease levels, disease categories, disease severity, or disease mechanism information.

[0051] For target nodes, the node embedding module can generate target node representations based on target name, target sequence features, target function, participating pathways, or target-related disease mechanisms.

[0052] Through the node embedding module, different types of nodes are converted into node representations of a unified dimension, enabling different types of nodes to participate in subsequent meta-path learning and hierarchical heterogeneous graph attention computation.

[0053] The meta-path learning module is used to generate candidate meta-paths from the heterogeneous graph, with patient nodes and their accompanying medication combinations as starting constraints and adverse event nodes as target constraints. The target meta-path is then determined based on the ability of the candidate meta-paths to represent the relationship between accompanying medication combinations and adverse events.

[0054] The meta-paths in this embodiment of the invention are not arbitrary graph paths, but rather formed around specific prediction objects in clinical trial monitoring. The prediction object is the probability of a potential association between the current patient and a certain adverse event standard concept node under their concomitant medication combination. Therefore, the meta-path learning module first determines the current patient node and at least two drug nodes connected to it as the initial path structure; simultaneously, it determines the adverse event standard concept node to be predicted in the global heterogeneous graph as the path target structure. Subsequently, the meta-path learning module generates multiple candidate meta-paths along the historical known association network and the current patient subgraph, considering medication relationships, disease relationships, drug interaction relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations.

[0055] like Figure 3 As shown, the heterogeneous graph of adverse drug events illustrates patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, as well as the relationships between different types of nodes. Figure 3 The node legend is used to distinguish different types of nodes, and the relationship legend is used to distinguish drug-drug relationships, disease relationships, drug-drug interaction relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations.

[0056] like Figure 3 As shown in the example of the target meta-path on the right, candidate meta-paths can include target meta-path 1, target meta-path 2, and target meta-path 3. Target meta-path 1 is a direct medication path connecting the patient node to the adverse event node via the drug node, representing the direct association between the current patient's use of a certain drug and the target adverse event. Target meta-path 2 is a drug interaction path connecting the patient node to the adverse event node via the drug node and another drug node, representing the influence of different drugs in the current patient's concomitant medication combination on the risk of the target adverse event through drug interactions. Target meta-path 3 is a disease-target association path connecting the patient node to the adverse event node via the disease node, target node, and drug node, representing the chain association between the current patient's disease background, drug target, drug node, and target adverse event. Figure 3 The thicker lines indicate paths that correspond to target path 1, target path 2, and target path 3, respectively.

[0057] After identifying candidate meta-paths, the meta-path learning module determines the target meta-path based on the contribution of each candidate meta-path to adverse event prediction. Contribution can be determined by the degree of improvement in prediction results after the introduction of the candidate meta-path, the importance of key nodes in the candidate meta-path, the strength of the association between the candidate meta-path and the target adverse event node, the length of the candidate meta-path, or the medical semantic completeness of the candidate meta-path. Candidate meta-paths with high contribution and that can effectively characterize the association between the concomitant medication combination and the adverse event are identified as target meta-paths.

[0058] In one implementation, adaptive meta-path learning can be performed by a path search agent. During model training, the path search agent performs path search in a heterogeneous graph containing historically known association networks and training patient subgraphs, and determines path rewards based on feedback from downstream adverse event prediction tasks. The path rewards are used to evaluate the improvement of candidate meta-paths on adverse event prediction performance and the path length of the candidate meta-paths. The state of the path search agent includes the current node representation, the target adverse event node representation, and the generated path representation. The actions of the path search agent include selecting the next neighbor node or the next neighbor node type from the current node. During model inference, the path search agent no longer calculates rewards in real-time based on the improvement in prediction performance of the current sample. Instead, based on the path search strategy obtained during training, it generates candidate meta-paths in the current patient subgraph and the global heterogeneous graph and determines the target meta-path.

[0059] In this way, the metapath learning module can avoid indiscriminate aggregation of the entire graph. Instead, it starts from the patient node and its accompanying medication combination, and forms a target metapath with predictive and explanatory significance around the target adverse event node.

[0060] like Figure 4 As shown, the hierarchical heterogeneous graph attention module is used to learn features of the target metapath and its associated nodes based on node-level attention, metapath-level attention and global attention, so as to obtain the risk representation of the patient under the combination of concomitant medications. Figure 4 The data processing order is shown for input, node-level attention, meta-path-level attention, global attention, risk prediction module, interpretable output module, and output results.

[0061] Figure 4 The input includes target meta-path 1, target meta-path 2, target meta-path 3, and a set of associated nodes for each target meta-path. The set of associated nodes may include drug nodes, disease nodes, target nodes, and adverse event nodes. Figure 4 The diagram below illustrates the node types corresponding to different node shapes.

[0062] Node-level attention is used to determine the importance of associated nodes in the same target metapath and output the importance or contribution of associated nodes. Specifically, node-level attention determines the importance of each associated node in the same target metapath based on the node representation of each associated node. For the same target metapath, if a drug node, disease node, or target node has a stronger association with the target adverse event node, or if its node representation contributes more to the prediction of adverse events, then that node receives a higher node contribution.

[0063] Meta-path-level attention is used to determine the importance of different target meta-paths and output the importance or contribution of each target meta-path. Specifically, meta-path-level attention compares the path representations formed by different target meta-paths to determine the degree of contribution of each target meta-path to the prediction of adverse events. For example, target meta-path 1, target meta-path 2, and target meta-path 3 correspond to different sources of risk, and meta-path-level attention is used to determine the relative importance of different target meta-paths in the current prediction task.

[0064] Global attention is used to combine patient node representation, concomitant medication combination representation, and the path contribution of different target meta-paths to determine the fusion weights of different target meta-paths in the current patient prediction task, and to generate a risk representation of the patient under the concomitant medication combination based on the fusion weights. Specifically, meta-path level attention focuses on determining the path contribution of each target meta-path itself to adverse event prediction, while global attention focuses on fusing different target meta-paths in the prediction context of the current patient and its concomitant medication combination, so that the final risk representation simultaneously includes patient individual characteristics, concomitant medication combination, disease background, drug target relationships, and adverse event association information.

[0065] The risk prediction module generates a risk representation vector based on the risk representation, and obtains the adverse event prediction probability and risk score based on the risk representation vector. The adverse event prediction probability characterizes the likelihood of a target adverse event occurring in the current patient with the given concomitant medication combination, and the risk score can be obtained by mapping the adverse event prediction probability.

[0066] The interpretable output module is used to determine key drug nodes, key drug combinations, target metapaths and their importance, risk causal tracing, and evidence chain display results based on node contribution, path contribution, and fusion weight. Figure 4 The output results are used to represent the final results that the system provides to clinical trial monitors, including risk scores, key drug nodes, key drug combinations, target metapaths, risk causal sources, and evidence chain presentation.

[0067] The risk prediction module is used to predict adverse events that may be induced by the combination of accompanying medications based on the risk representation and to generate a risk score.

[0068] In one implementation, the risk prediction module inputs the risk representation output by the hierarchical heterogeneous graph attention module into the prediction layer. Based on the matching relationship between the risk representation and each adverse event node, the prediction layer obtains the predicted probability of the target adverse event or multiple candidate adverse events. This predicted probability is used to characterize the likelihood of the patient experiencing the corresponding adverse event under the current concomitant medication combination.

[0069] The predicted probability of adverse events can be obtained through binary classification, multi-class classification, or multi-label prediction. When the system predicts a specific adverse event, the risk prediction module outputs the probability of occurrence corresponding to that adverse event. When the system predicts multiple adverse events, the risk prediction module outputs the probability of occurrence for each adverse event and determines the adverse events requiring priority attention based on the probability magnitude.

[0070] The risk score is generated based on the predicted probability of adverse events. This predicted probability is derived from a risk prediction module based on a risk representation, reflecting the likelihood of a patient experiencing a target adverse event given the current combination of medications. The risk score can be represented in the range of 0 to 100 and can be mapped from the predicted probability of adverse events. A higher predicted probability corresponds to a higher risk score. The system can classify risk levels based on the risk score, such as low risk, medium risk, or high risk.

[0071] The contribution of key drug nodes is determined by node-level attention. For drug nodes in different target meta-paths, the system determines their impact on adverse event prediction results based on their contribution in the corresponding target meta-path. If the same drug node appears in multiple target meta-paths, the system can aggregate the node contributions of the drug node in different target meta-paths and determine the overall contribution of the drug node based on the aggregated results.

[0072] The contribution of key drug combinations is determined jointly by multiple drug nodes in the drug interaction path. When a target metapath contains at least two drug nodes with drug interactions, the system determines the contribution of the drug combination to the adverse event prediction results based on the path contribution and the node contribution of each drug node in the path. If the same drug combination appears in multiple target metapaths, the system can combine the contribution results from different paths to rank the drug combinations.

[0073] The contribution of a target meta-path is determined jointly by meta-path-level attention and global attention. Meta-path-level attention is used to assess the contribution of each target meta-path to adverse event prediction, while global attention is used to determine the fusion weight of different target meta-paths in the current patient prediction task. The system can determine the contribution of a target meta-path based on its importance and fusion weight.

[0074] The contribution of key drug nodes, key drug combinations, and target metapaths are not directly used as the sum of the probability of risk occurrence, but are used to explain the source of the risk score and to generate the results of risk cause tracing and evidence chain display.

[0075] The interpretable output module is used to determine key drug nodes, key drug combinations and corresponding target meta-paths based on the importance of associated nodes, the importance of target meta-paths and the fusion weights of different target meta-paths, and output interpretable prediction results including risk cause tracing and evidence chain display.

[0076] Explainable prediction results can include predictions of adverse events, risk scores, key drug nodes, key drug combinations, target metapaths, importance of associated nodes, importance of target metapaths, and evidence chain presentation results.

[0077] Key drug nodes can be identified based on node-level attention. The system can rank drug nodes according to their contribution and identify those with higher contributions as key drug nodes.

[0078] Key drug combinations can be determined based on drug interaction pathways. For a target metapath containing at least two drug nodes, the system determines the importance of the drug combination based on the path contribution and the node contribution of each drug node in the path, and outputs the drug combination with the higher contribution as the key drug combination.

[0079] The target meta-path can be determined based on meta-path level attention and global attention. The system can sort the target meta-paths according to their contribution and output the target meta-paths with higher contributions as the interpretation results.

[0080] The evidence chain display results show at least some of the nodes and their corresponding contributions from the patient node, drug node, disease node, target node, and adverse event node in the order of the target meta-path. For example, the system can display a direct drug use evidence chain of "patient node - drug node - adverse event node", a drug interaction evidence chain of "patient node - drug node - drug node - adverse event node", and a disease-target association evidence chain of "patient node - disease node - target node - drug node - adverse event node".

[0081] Risk triggers can be identified by combining key drug nodes, key drug combinations, and target meta-pathways. When a risk score is high, the system can further display the main contributing drugs, main drug combinations, main target meta-pathways, and the contribution of each node corresponding to that risk score, enabling clinical trial monitors to understand the source of the risk score.

[0082] Through the interpretable output module, the system not only outputs the risk prediction results, but also the node source, path source, and contribution source corresponding to the risk results, thereby improving the interpretability of the risk prediction results of adverse drug events during clinical trial monitoring.

[0083] like Figure 1 As shown, the method for predicting the association of adverse events with concomitant medication for clinical trial monitoring provided in this embodiment of the invention includes steps S101 to S106.

[0084] Step S101: Acquire data and perform standardization processing Acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data from clinical trials, and standardize drug names, adverse event terminology, and disease information.

[0085] In this step, data can be obtained from clinical trial data management systems, electronic data capture systems, drug databases, medical terminology databases, drug target databases, or historical clinical trial databases. Patient data may include age, gender, underlying diseases, comorbidities, trial group, previous treatment history, and laboratory test results. Concomitant medication data may include drug name, drug class, dosage, frequency of administration, route of administration, and medication records. Adverse event data may include the name, severity, time of occurrence, outcome, and management of the adverse event. Disease data may include disease diagnosis, disease code, underlying diseases, and comorbidities. Drug target data may include the drug's target, target function, and information on target-related diseases or pathways.

[0086] After standardization, drug names, adverse event terms, and disease names are unified to their corresponding standard expressions and mapped to drug nodes, adverse event nodes, and disease nodes, respectively. Patient data and drug target data are mapped to patient nodes and target nodes, respectively.

[0087] Step S102: Construct a heterogeneous graph of concomitant adverse drug events A heterogeneous graph of adverse drug events is constructed based on standardized data. This heterogeneous graph includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-drug interaction relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations.

[0088] like Figure 3As shown, patient nodes and drug nodes are connected through medication relationships, representing a patient's use of a particular drug. Patient nodes and disease nodes are connected through disease relationships, representing the presence of a patient's underlying disease or comorbidity. Different drug nodes are connected through drug interaction relationships, representing potential interactions between different drugs in a combination of medications. Drug nodes and target nodes are connected through drug-target relationships, representing the drug's target. Target nodes and disease nodes are connected through target-disease relationships, representing the association between the target and the disease mechanism. Drug nodes and adverse event nodes are connected through drug-adverse event association relationships, representing known, suspected, or historical statistical associations between drugs and adverse events.

[0089] In step S102, the patient, concomitant medication combination, adverse events, disease, and drug target are unified into the same graph structure, thereby providing a foundation for subsequent path learning and attention calculation.

[0090] Step S103: Perform node embedding By embedding different types of nodes in a heterogeneous graph, a node representation in a unified vector space is obtained.

[0091] In this step, patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes are each generated with an initial node representation using a feature extraction method that matches their node type. These initial node representations for different node types are further mapped to a vector space of the same dimension, enabling different node types to participate in subsequent meta-path learning and hierarchical heterogeneous graph attention computation.

[0092] In one implementation, the patient node representation is generated from the patient's clinical characteristics, the drug node representation is generated from the drug's chemical structure and pharmacological characteristics, the adverse event node representation is generated from the semantics of adverse event terminology, the disease node representation is generated from the disease coding hierarchy, and the target node representation is generated from the target sequence characteristics or functional characteristics.

[0093] Step S104: Generate candidate meta paths and determine the target meta path Using patient nodes and their accompanying medication combinations as initial constraints and adverse event nodes as target constraints, candidate meta-paths are generated from heterogeneous graphs and target meta-paths are determined based on adaptive meta-path learning. Target meta-paths represent the associations between accompanying medication combinations and adverse events formed through disease nodes, target nodes, or drug interactions.

[0094] like Figure 3As shown, the target meta-path can include target meta-path 1, target meta-path 2, and target meta-path 3. Target meta-path 1 can be represented as a direct medication path connecting the patient node to the adverse event node via the drug node; target meta-path 2 can be represented as a drug interaction path connecting the patient node to the adverse event node via at least two drug nodes with drug interaction relationships; and target meta-path 3 can be represented as a disease-target association path connecting the patient node to the adverse event node via the disease node, target node, and drug node.

[0095] In one implementation, adaptive meta-path learning is performed by a path search agent. During the training phase, the path search agent determines the path search state based on the current node representation, the target adverse event node representation, and the generated path representation; determines the available actions based on the current node's neighbor nodes or neighbor node types; and determines the path reward based on the downstream adverse event prediction performance improvement and the path length penalty. During the inference phase, the path search agent determines the target meta-path from the candidate meta-paths based on the trained path search strategy.

[0096] Through step S104, the model can start from the patient and their combination of medications, and determine the target meta-path that contributes to the prediction around the target adverse event node, thereby avoiding the indiscriminate aggregation of full graph information.

[0097] Step S105: Perform hierarchical heterogeneous graph attention feature learning Based on hierarchical heterogeneous graph attention network, feature learning is performed on the target metapath and its associated nodes to obtain the risk representation of patients under the combination of concomitant medications.

[0098] like Figure 4 As shown, the hierarchical heterogeneous graph attention network includes node-level attention, meta-path-level attention, and global attention. Node-level attention is used to determine the importance or contribution of associated nodes based on the set of associated nodes in each target meta-path; meta-path-level attention is used to determine the importance or contribution of different target meta-paths based on the path representations of target meta-path 1, target meta-path 2, and target meta-path 3; global attention is used to combine patient node representation, concomitant medication combination representation, and the path contribution of different target meta-paths to determine the fusion weight of different target meta-paths in the current patient prediction task, and generate a risk representation of the patient under the concomitant medication combination. Through step S105, patient individual characteristics, concomitant medication combination, disease background, target relationship, and adverse event association information are collaboratively fused to form a risk representation of the patient under the concomitant medication combination.

[0099] Step S106: Generate risk scores and interpretable prediction results Based on the risk representation, adverse events that may be induced by the combination of medications are predicted and a risk score is generated. Furthermore, based on the importance of associated nodes, the importance of target meta-paths, and the fusion weight of different target meta-paths, key drug nodes, key drug combinations, and corresponding target meta-paths are determined, and interpretable prediction results including risk cause tracing and evidence chain presentation are output.

[0100] Risk scores can be used to represent the degree of risk a patient faces in developing a target adverse event under the current combination of concomitant medications. Risk scores can be generated based on the predicted probability of adverse events. The contribution of key drug nodes, key drug combinations, and target metapaths are used to explain the source of the risk score and to generate results for risk causal attribution and evidence chain presentation.

[0101] Explainable prediction results can include predicted adverse events, risk scores, key drug nodes, key drug combinations, target metapaths, importance of associated nodes, importance of target metapaths, and evidence chain presentation results. The evidence chain presentation results can display at least some nodes from the patient node, drug node, disease node, target node, and adverse event node, along with their corresponding contributions, in the node order of the target metapath.

[0102] Through step S106, the system can not only output adverse event risk prediction results, but also explain the source of the graph structure of the risk score and the risk trigger path, thereby improving the interpretability of risk prediction results in clinical trial monitoring scenarios.

[0103] The method for predicting the association of adverse events with accompanying medication of the present invention has the following application embodiments: Example 1: Prediction of Adverse Events from Concomitant Medication Use in Oncology Clinical Trials In an oncology clinical trial, participants received treatment with the investigational drug and chemotherapy, along with supportive care. The system acquired the participant's patient data, concomitant medication data, adverse event data, disease data, and drug target data.

[0104] Patient data includes age, sex, tumor type, disease stage, previous treatment history, and laboratory test results. Concomitant medication data includes the investigational drug, chemotherapy drugs, and supportive care drugs. Adverse event data includes hematologic adverse events, gastrointestinal adverse events, fatigue-related or immune-related adverse events, etc. Disease data includes tumor diagnosis and comorbidities. Drug target data includes the corresponding targets of the investigational drug and chemotherapy drugs.

[0105] The system first standardizes drug names, adverse event terminology, and disease information, and then constructs a heterogeneous graph of concomitant adverse events. In this graph, patient nodes are connected to multiple drug nodes, drug interactions exist between drug nodes, drug nodes are connected to target nodes and adverse event nodes, and target nodes are connected to disease nodes.

[0106] The meta-path learning module uses the patient node and its accompanying medication combination as initial constraints and the target adverse event node as target constraints to generate candidate meta-paths. Examples include a direct medication path connecting the patient node to the neutropenia node via a chemotherapy drug node; a drug interaction path connecting the patient node to the thrombocytopenia node via two chemotherapy drug nodes; and a disease-target association path connecting the patient node to the immune-related adverse event node via a disease node, target node, and drug node.

[0107] The hierarchical heterogeneous graph attention module performs feature learning on the aforementioned target meta-paths and their associated nodes to identify key drug nodes, key drug combinations, and target meta-paths. The risk prediction module outputs adverse event prediction results and risk scores based on the risk representation. The interpretable output module further outputs risk causation tracing and evidence chain presentation results, enabling clinical trial monitors to identify whether the corresponding adverse event risk may originate from a key drug, a drug combination, or a disease-target association path.

[0108] Example 2: Risk monitoring in a multicenter clinical trial In multicenter clinical trials, subjects from different research centers may have different disease backgrounds, concomitant medication combinations, and adverse event records. The system in this embodiment of the invention can receive standardized clinical trial data uploaded by multiple research centers and construct or update the corresponding local structure of the concomitant medication adverse event heterogeneity graph for each patient node.

[0109] For each patient, the system uses the patient node and its accompanying medication combination as initial constraints, and the candidate adverse event node as the target constraint to generate candidate meta-paths and determine the target meta-path. Subsequently, the hierarchical heterogeneous graph attention module calculates the contribution of different nodes and different target meta-paths, the risk prediction module generates a risk score, and the interpretable output module outputs the results of key drug nodes, key drug combinations, and evidence chains.

[0110] When a patient's risk score reaches a preset risk level, the system can generate a risk alert for further review by clinical trial monitors. This risk alert may include predicted adverse events, risk score, key drug combinations, target pathways, and evidence chain presentation results.

[0111] Example 3: Risk Assessment of Concomitant Medication Use in Special Populations For elderly patients, patients with abnormal liver or kidney function, or subjects with multiple comorbidities, the patient's disease background and individual characteristics may significantly affect the risk of adverse drug events. In this embodiment, the system can incorporate information such as age, liver and kidney function, comorbidities, and past medication history as part of the patient node features into the node embedding.

[0112] When patients have underlying conditions such as chronic kidney disease, hypertension, or diabetes, and are simultaneously using investigational drugs and multiple medications for their underlying conditions, the system can establish relationships between patient nodes, disease nodes, drug nodes, target nodes, and adverse event nodes in a heterogeneous graph of concomitant adverse events. The meta-path learning module can identify the associated paths between the patient's disease background, drug targets, and adverse events. The hierarchical heterogeneous graph attention module can determine the contribution of disease nodes, drug nodes, or target nodes to the target adverse event.

[0113] Through the above processing, the system can output a patient's adverse event risk score under a combination of concomitant medications and provide a corresponding chain of evidence. This chain of evidence can be used to explain how the association between a certain combination of concomitant medications, the patient's disease background, and the drug target affects the risk of the target adverse event.

[0114] Example 4: Risk Prediction of New Drug Combinations In cases where sufficient historical adverse event data is lacking for new drug combinations, the system of this invention can use the chemical structural features, pharmacological features, drug target data, and existing drug-adverse event correlations of drug nodes to predict risks.

[0115] Once a new drug combination enters clinical trials, the system acquires the drug node representations within that combination and generates candidate metapaths based on drug-target relationships, target-disease relationships, and drug-adverse event associations. For drug combinations lacking sufficient historical data, the system can still generate risk representations using drug targets, disease mechanisms, and similar drug adverse event association information.

[0116] The meta-path learning module identifies the target meta-path from candidate meta-paths, the hierarchical heterogeneous graph attention module generates a risk representation for the patient under this new drug combination based on node contribution and path contribution, the risk prediction module generates a risk score, and the interpretable output module outputs the results of risk cause tracing and evidence chain presentation. Therefore, the system can provide auxiliary analytical results for monitoring the risk of concomitant adverse events associated with new drug combinations in the early stages of clinical trials.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the association of concomitant adverse events with medication in clinical trial monitoring, characterized in that, include: Acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data from clinical trials, and standardize drug names, adverse event terminology, and disease information; A heterogeneous graph of adverse drug events is constructed based on standardized data. The heterogeneous graph includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-target relationships, target-disease relationships, and drug-adverse event associations. Node embedding is performed on different types of nodes in the heterogeneous graph to obtain node representations in a unified vector space; Using patient nodes and their accompanying medication combinations as initial constraints and adverse event nodes as target constraints, candidate meta-paths are generated from the heterogeneous graph based on adaptive meta-path learning, and target meta-paths are determined. The target meta-paths represent the association between the accompanying medication combinations and adverse events through disease nodes, target nodes, or drug interactions. Based on a hierarchical heterogeneous graph attention network, feature learning is performed on the target meta-path and its associated nodes to obtain a risk representation of the patient under the combination of concomitant medications. Based on the risk representation, predict the adverse events that the accompanying medication combination may induce and generate a risk score; The hierarchical heterogeneous graph attention network determines the importance of associated nodes, the importance of target meta-paths, and the fusion weights of different target meta-paths based on node-level attention, meta-path-level attention, and global attention, respectively. Based on the importance of associated nodes, the importance of target meta-paths, and the fusion weights, it determines key drug nodes, key drug combinations, and corresponding target meta-paths, and outputs interpretable prediction results including risk causal tracing and evidence chain display.

2. The method for predicting the association of adverse events related to medication use in clinical trial monitoring according to claim 1, characterized in that, The standardization process includes: standardizing drug names in the accompanying medication data into drug codes, standardizing adverse event terms in the adverse event data into standard adverse event terms, standardizing disease names in the disease data into disease codes, and mapping the patient data, drug codes, standard adverse event terms, disease codes, and drug target data to corresponding patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, respectively.

3. The method for predicting the association of adverse events related to medication use in clinical trial monitoring according to claim 1, characterized in that, The patient node is used to characterize individual patient characteristics, the drug node is used to characterize investigational drugs and concomitant drugs, the adverse event node is used to characterize historically recorded adverse events or concepts of adverse events to be predicted, the disease node is used to characterize underlying diseases or comorbidities, and the target node is used to characterize drug targets. The medication relationship connects patient nodes and drug nodes, the disease relationship connects patient nodes and disease nodes, the drug interaction relationship connects different drug nodes, the drug-target relationship connects drug nodes and target nodes, the target-disease relationship connects target nodes and disease nodes, and the drug-adverse event association relationship connects drug nodes and adverse event nodes.

4. The method for predicting the association of concomitant adverse events with medication for clinical trial monitoring according to claim 1, characterized in that, The node embedding includes: generating patient node representations based on patient clinical characteristics, generating drug node representations based on drug chemical structure features and pharmacological features, generating adverse event node representations based on adverse event terminology semantics, generating disease node representations based on disease coding hierarchy, and generating target node representations based on target sequence features or functional features, so that the node representations of different types of nodes are in the same dimension vector space.

5. The method for predicting the association of concomitant adverse events with medication for clinical trial monitoring according to claim 1, characterized in that, The adaptive meta-path learning includes: using the current patient node and at least two drug nodes connected to it as the starting structure of the path, using the concept node of the adverse event to be predicted in the global heterogeneous graph as the target structure of the path, generating candidate meta-paths based on the drug relationship, disease relationship, drug interaction relationship, drug-target relationship, target-disease relationship and drug-adverse event association relationship in the heterogeneous graph, and determining the target meta-path according to the contribution of each candidate meta-path to the prediction of adverse events.

6. The method for predicting the association of concomitant adverse events in clinical trial monitoring according to claim 5, characterized in that, The adaptive meta-path learning is performed by a path search agent. During the training phase, the state of the path search agent includes the current node representation, the target adverse event node representation, and the generated path representation. The actions of the path search agent include selecting the next neighbor node or the next neighbor node type from the current node. The reward of the path search agent is determined based on the downstream adverse event prediction performance improvement and the path length penalty term. During the inference phase, the path search agent determines the target meta-path based on the path search strategy obtained during training.

7. The method for predicting the association of adverse events related to medication use in clinical trial monitoring according to claim 1, characterized in that, The target metapath includes at least two of the following: a direct medication path where the patient node is connected to the adverse event node via a drug node; a drug interaction path where the patient node is connected to the adverse event node via at least two drug nodes with drug interaction relationships; and a disease-target association path where the patient node is connected to the adverse event node via a disease node, a target node, and a drug node.

8. The method for predicting the association of adverse events related to medication use in clinical trial monitoring according to claim 1, characterized in that, The node-level attention is used to determine the node contribution of drug nodes, disease nodes, target nodes, and adverse event nodes in the same target metapath to adverse event prediction. The metapath-level attention is used to determine the path contribution of different target metapaths to adverse event prediction. The global attention is used to combine the patient node representation, the concomitant medication combination representation, and the path contribution to determine the fusion weight of different target metapaths in the current patient prediction task, so as to generate the risk representation of the patient under the concomitant medication combination.

9. The method for predicting the association of concomitant adverse events with medication for clinical trial monitoring according to claim 1, characterized in that, The risk score is generated based on the predicted probability of adverse events; the interpretable prediction results include predicted adverse events, risk scores, key drug nodes, key drug combinations, target metapaths, importance of associated nodes, importance of target metapaths, and evidence chain presentation results. The key drug nodes are determined based on node contribution, the key drug combinations are determined based on the contribution of drug nodes and path in the drug interaction path, and the evidence chain presentation results display at least some of the nodes and their corresponding contributions from patient nodes, drug nodes, disease nodes, target nodes, and adverse event nodes in the node order of the target metapath.

10. A predictive system for concomitant adverse events related to medication use in clinical trial monitoring, characterized in that, include: The data standardization module is used to acquire patient data, concomitant medication data, adverse event data, disease data, and drug target data in clinical trials, and to standardize drug names, adverse event terms, and disease information. The heterogeneous graph construction module is used to construct a heterogeneous graph of accompanying adverse drug events based on standardized data. The heterogeneous graph includes patient nodes, drug nodes, adverse event nodes, disease nodes, and target nodes, and includes drug-drug relationships, disease relationships, drug-drug interaction relationships, drug-target relationships, target-disease relationships, and drug-adverse event association relationships. The node embedding module is used to embed different types of nodes in the heterogeneous graph to obtain node representations in a unified vector space. The meta-path learning module is used to generate candidate meta-paths and determine target meta-paths from the heterogeneous graph, using the current patient node and at least two drug nodes connected to it as the path starting structure and the adverse event concept node to be predicted in the global heterogeneous graph as the path target structure. The hierarchical heterogeneous graph attention module is used to learn features of the target meta-path and its associated nodes based on node-level attention, meta-path-level attention and global attention, so as to obtain the risk representation of the patient under the combination of concomitant medications. The risk prediction module is used to obtain the predicted probability of adverse events based on the risk representation, and to generate a risk score result based on the predicted probability of adverse events. The interpretable output module is used to determine key drug nodes, key drug combinations and corresponding target meta-paths based on the importance of associated nodes, the importance of target meta-paths and the fusion weight, and output interpretable prediction results including risk cause tracing and evidence chain display.

Images