Data collection and drug safety signal mining methods and agents for PMS

Abstract

The present invention provides a method and intelligent agent for data collection and drug safety signal mining for PMS, which relates to the field of drug monitoring technology, including standardizing PMS data using a medical text pre-training model, and integrating drug molecular characteristics and clinical medication information to construct a multi-layer drug knowledge graph. A heterogeneous graph neural network with a time series feature extraction function is used to represent the graph, and the drug action mechanism rules are integrated into the training process to obtain a node vector representation of multi-dimensional information fusion. Based on the node vector representation, a multi-task learning framework is constructed to simultaneously complete the drug-adverse reaction association probability prediction and risk level prediction, and combine time series analysis to detect dynamic change characteristics to generate candidate signals. Finally, a hierarchical evaluation system is constructed to verify the candidate signal, and its reliability is confirmed in combination with clinical data, the target signal is output, the warning level is determined, and differentiated monitoring cycles and trigger thresholds are set to perform real-time risk warnings.

Description

Technical Field

[0001] The present invention relates to drug monitoring technology, and in particular to a PMS-oriented data collection and drug safety signal mining method and an intelligent agent. Background Art

[0002] Post-Marketing Surveillance (PMS) aims to timely detect potential adverse reactions and safety risks and ensure the safety of public drug use by continuously collecting and analyzing the safety data of drugs in actual use. Traditional PMS systems mainly rely on spontaneous reporting systems and expert reviews, and have limitations such as insufficient data collection, low analysis efficiency, and long response time, making it difficult to detect potential drug risks in a timely manner. In order to improve the intelligence and real-time performance of PMS, PMS systems based on big data and artificial intelligence have gradually been studied and applied in recent years.

[0003] The present invention proposes a method for data collection and drug safety signal mining for PMS. The drug names and adverse reaction descriptions are standardized by using a medical text pre-training model to construct standardized drug adverse reaction data. At the same time, the drug molecular feature information and clinical medication information are integrated to establish a multi-layer drug knowledge graph containing drug and adverse reaction nodes, and a heterogeneous graph neural network is used for representation learning to achieve accurate identification of drug adverse reaction signals and dynamic risk prediction. In addition, a hierarchical early warning model is constructed, and differentiated early warning is achieved in combination with expert rules, which significantly improves the monitoring efficiency and signal recognition accuracy of the PMS system. Summary of the invention

[0004] The embodiment of the present invention provides a PMS-oriented data collection and drug safety signal mining method and intelligent agent, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] A method for data collection and drug safety signal mining for PMS is provided, including:

[0007] Acquire historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping on drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by integrating drug molecular feature information and clinical medication information, wherein the multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes;

[0008] A heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, and training samples are constructed through a node type-aware sampling strategy. The drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network to obtain a node vector representation of multidimensional information fusion. Based on the node vector representation, a multi-task learning framework is constructed. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, and combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification;

[0009] Construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized adverse drug reaction data.

[0010] In an optional embodiment,

[0011] The medical text pre-training model is used to standardize the mapping of drug names and adverse reaction descriptions in the original data to generate standardized adverse drug reaction data. Based on the standardized adverse drug reaction data, the drug molecular feature information and clinical medication information are integrated to construct a multi-layer drug knowledge graph, including:

[0012] Inputting the original adverse drug reaction data into a medical text pre-training model, the medical text pre-training model performs word segmentation on the input text and extracts word vectors and term vectors, concatenating the word vectors with the term vectors to form a feature vector, inputting the feature vector into a bidirectional long short-term memory unit to extract context features, and performing sequence labeling through a conditional random field based on the context features to obtain initial drug entities and initial adverse reaction entities;

[0013] Calculating the string similarity and semantic similarity between the initial drug entity and the standard medical term, determining the initial drug entity whose weighted combination value of the two similarities exceeds the mapping threshold as the standard drug entity, and mapping the initial adverse reaction entity to the standard adverse reaction entity to generate standardized drug adverse reaction data;

[0014] The simplified linear input canonical representation of the drug molecule is converted into a molecular structure diagram with atom type and chemical bond type annotations. The molecular structure diagram is subjected to local feature extraction through multi-layer spatial convolution, and the extracted local features are aggregated through attention weights to obtain a drug molecule feature vector;

[0015] Input clinical medication text into the bidirectional attention network, calculate the attention weight with standard medical terms, and obtain standardized clinical medication entities;

[0016] The sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; the molecular similarity is calculated using drug molecular feature vectors to construct a molecular structure association network; and a clinical association network is constructed based on standardized clinical drug entities;

[0017] Input the temporal association network, the molecular structure association network and the clinical association network into a hierarchical graph attention network, generate a node vector representation through a random walk algorithm, the node vector representation obtains multi-layer network features through a graph attention layer, and the multi-layer network features are input into a recurrent neural network to extract temporal features;

[0018] The multi-layer network features and time series features are input into a heterogeneous graph neural network, and a multi-layer drug knowledge graph containing node vector representations is obtained through feature transformation training to form a multi-dimensional association network of drug information nodes and adverse reaction information nodes.

[0019] In an optional embodiment,

[0020] The sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; the molecular similarity is calculated using the drug molecular feature vector to construct a molecular structure association network; the clinical association network is constructed based on the standardized clinical drug entities, including:

[0021] A sliding time window is used to detect the co-occurrence relationship between the standard drug entity and the standard adverse reaction entity in the time dimension to form a drug-adverse reaction entity pair, and a temporal co-occurrence metric value of the drug-adverse reaction entity pair is calculated for each time window. The temporal co-occurrence metric value is determined by the ratio of the number of co-occurrences of the drug-adverse reaction entity pair in the window to the total number of occurrences of each drug-adverse reaction entity pair;

[0022] Performing a temporal significance test on the drug-adverse reaction entity pairs, generating a co-occurrence probability distribution by a random permutation method, establishing connections between the drug-adverse reaction entity pairs that pass the significance test, and using the temporal co-occurrence measurement value as a connection weight value to construct a temporal association network;

[0023] Inputting the drug molecule feature vector into a similarity calculation model, the similarity calculation model performs attention weighted calculation on the chemical substructures in the drug molecule feature vector, outputs a molecular similarity matrix, inputs the molecular similarity matrix into a spectral clustering model for drug grouping, and connects drug molecules with similarities higher than a preset threshold according to the grouping results to construct a molecular structure association network;

[0024] Forming a drug combination sequence according to the clinical drug entity, calculating the co-occurrence times of drugs in the drug combination sequence, obtaining the clinical co-occurrence support of the drug combination, establishing a Markov chain model for the drug combination sequence in time sequence, and calculating the transition probability between drugs in adjacent time sequence based on the Markov chain model;

[0025] Constructing a bipartite graph of drugs and clinical indications, calculating the drug-clinical indication co-occurrence probability, calculating the mutual information ratio of the drug edge probability and the clinical indication edge probability according to the drug-clinical indication co-occurrence probability to obtain the drug-clinical indication association strength, and weightedly combining the clinical co-occurrence support, drug transfer probability and drug-clinical indication association strength as the connection weight value to construct a clinical association network;

[0026] The connection weights in the temporal association network, the connection weights in the molecular structure association network, and the connection weights in the clinical association network are fused to construct a multi-layer drug association network.

[0027] In an optional embodiment,

[0028] A heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, a node type-aware sampling strategy is used to construct training samples, and the drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network to obtain a node vector representation of multidimensional information fusion, including:

[0029] Receiving a multidimensional association network and a drug action mechanism rule, performing a linear transformation on a node feature vector in the multidimensional association network, and generating a query vector, a key vector, and a value vector;

[0030] An attention score matrix is ​​obtained by performing a dot product operation on the query vector and the key vector, the attention score matrix is ​​normalized to obtain a temporal attention weight, a matrix multiplication operation is performed on the temporal attention weight and the value vector to obtain a temporal context vector, and the temporal context vector is concatenated with the node feature vector and then the temporal enhancement feature is obtained through a feedforward neural network;

[0031] The time series enhancement feature is weightedly summed with the edge weight in the multidimensional association network to obtain the edge weight of the fused time series feature, the degree centrality of each node is calculated as the node importance score according to the edge weight of the fused time series feature, the nodes are divided into different categories according to the node type, and probability sampling is performed in each category according to the node importance score to obtain a training node set;

[0032] Extracting drug nodes associated with the mechanism of action from the drug action mechanism rules as drug node pairs, calculating the cosine similarity of the drug node pairs in the training node set, and taking the difference between the similarity and a preset node similarity threshold as the rule loss;

[0033] Input the training node set into the heterogeneous graph neural network for node classification, use cross entropy to calculate the classification loss, and use the weighted sum of the classification loss and the rule loss as the total loss function;

[0034] Based on the total loss function, the parameters of the heterogeneous graph neural network are gradient updated, and the iteration is repeated until the preset number of iterations is met to obtain the final node vector representation of multi-dimensional information fusion.

[0035] In an optional embodiment,

[0036] Based on the node vector representation, a multi-task learning framework is constructed. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification, including:

[0037] Obtaining and concatenating the drug node vector and the adverse reaction node vector in the node vector representation, extracting features through a multi-layer attention-enhanced feedforward neural network with residual connections to generate a drug-adverse reaction shared feature vector;

[0038] Constructing a time series feature sequence matrix, wherein the rows of the time series feature sequence matrix represent different time windows, and the columns represent feature sets corresponding to the drug node vector and the adverse reaction node vector in each time window, wherein the feature set includes statistical association features calculated based on an adaptive time window, drug molecular structure features, and drug target features;

[0039] Inputting the temporal feature sequence matrix into a bidirectional long short-term memory network provided with a gating mechanism, wherein the gating mechanism adaptively adjusts the feature weight according to the importance of the time window to generate a temporal context vector;

[0040] A multi-head attention mechanism is used to fuse the drug-adverse reaction shared feature vector with the temporal context vector to obtain a drug-adverse reaction multi-task feature vector, and the drug-adverse reaction multi-task feature vector is input into an association prediction network and a risk prediction network respectively to generate an initial drug-adverse reaction association probability and an initial risk level probability;

[0041] Calculate the binary cross entropy loss based on the initial drug-adverse reaction association probability, calculate the multi-category cross entropy loss based on the initial risk level probability, construct a multi-task loss function in combination with the time series smoothing constraint, update the model parameters by iteratively optimizing the multi-task loss function, and obtain the optimized drug-adverse reaction association probability and risk level probability;

[0042] The kernel density estimation method is used to calculate the probability distribution of the current time window based on the optimized drug-adverse reaction association probability, determine the dynamic association threshold, and screen candidate drug safety signals from the optimized drug-adverse reaction association probability and risk level probability according to the dynamic association threshold and the preset risk threshold.

[0043] In an optional embodiment,

[0044] Constructing a hierarchical evaluation system including a signal strength dimension and a clinical relevance dimension, verifying the candidate drug safety signal based on the hierarchical evaluation system, confirming the reliability of the candidate drug safety signal in combination with clinical data, outputting a target drug safety signal, mapping the target drug safety signal to the hierarchical evaluation system, determining an early warning level based on a mapping position, setting a differentiated monitoring period and a trigger threshold according to the early warning level, and inputting the node vector representation into an early warning analysis model, and performing real-time risk early warning on the standardized adverse drug reaction data, including:

[0045] Construct a hierarchical evaluation system of signal strength dimension and clinical relevance dimension. The signal strength dimension is based on the strength of drug-adverse reaction causal relationship, the number of adverse reaction reports, and the time trend of adverse reactions to construct a weighted scoring model to obtain a signal strength score. The clinical relevance dimension is based on the severity of adverse reactions, population sensitivity, and clinical interventionability to construct a hierarchical analysis network to obtain a clinical relevance score.

[0046] Inputting the candidate drug safety signal, signal strength score, and clinical relevance score into the screening model to obtain a first screening result, inputting the first screening result into a cohort analysis unit, a literature analysis unit, and a social media analysis unit, respectively, and combining the verification results of each unit using an integrated learning model to obtain a multi-source verification result;

[0047] Inputting the drug molecular characteristics, metabolic information, action network information and the multi-source verification results corresponding to the candidate drug safety signal into a graph neural network model, and determining the target drug safety signal based on the output results of the graph neural network model;

[0048] Inputting the target drug safety signal, the signal strength score, and the clinical relevance score into a pre-trained deep reinforcement learning model to obtain a warning level;

[0049] A monitoring strategy is constructed according to the warning level, and the monitoring strategy and the time series characteristics of the target drug safety signal are input into a hybrid neural network model. The time series characteristics include the number characteristics of adverse reaction reports, the cumulative proportion characteristics and the geographical distribution characteristics. Based on the output results of the hybrid neural network model, risk warnings are issued for adverse drug reaction data.

[0050] In an optional embodiment,

[0051] The target drug safety signal, the signal strength score and the clinical relevance score are input into a pre-trained deep reinforcement learning model to obtain the warning levels including:

[0052] Collect characteristic information of the target drug safety signal, the characteristic information including signal structured representation, signal intensity temporal variation characteristics, clinical relevance score characteristics and historical warning records, construct the characteristic information into a state space and map it into a state vector;

[0053] Constructing an action space for warning level adjustment, the action space including a warning level raising action, a warning level lowering action and a warning level maintaining action, mapping the action space into an action vector, and the action vector and the state vector form a state-action pair;

[0054] Constructing a reward function for early warning evaluation, calculating a weighted reward value based on an early warning accuracy index, an early warning timeliness index, and a supervision resource consumption index, and forming a training sample with the state-action pair and the weighted reward value;

[0055] Constructing a dual learning network structure including a current network and a target network, storing the training samples in an experience replay pool, sampling from the experience replay pool to obtain a state transition sequence, inputting the state transition sequence into the dual learning network structure, and using a temporal difference algorithm to alternately optimize and update the parameters of the current network and the target network;

[0056] A decreasing greedy strategy is used for action selection. The corresponding action is selected based on the current state and environmental feedback is obtained. The environmental feedback is input into the dual learning network structure to obtain the optimal early warning decision strategy. The real-time feature vector of the target drug safety signal is input into the optimal early warning decision strategy, and the corresponding early warning level is output.

[0057] According to a second aspect of the embodiments of the present invention,

[0058] Provide a data collection and drug safety signal mining agent for PMS, including:

[0059] The first unit is used to obtain historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping of drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by fusing drug molecular feature information and clinical medication information, wherein the multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes;

[0060] The second unit is used to use a heterogeneous graph neural network with a time series feature extraction function to perform representation learning on the multidimensional association network, construct training samples through a node type-aware sampling strategy, integrate drug action mechanism rules into the training process of the heterogeneous graph neural network, obtain a node vector representation of multidimensional information fusion, and construct a multi-task learning framework based on the node vector representation. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification;

[0061] The third unit is used to construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized drug adverse reaction data.

[0062] According to a third aspect of the embodiments of the present invention,

[0063] An electronic device is provided, comprising:

[0064] processor;

[0065] a memory for storing processor-executable instructions;

[0066] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0067] A fourth aspect of the embodiments of the present invention is:

[0068] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0069] In this embodiment, the original data is standardized and structured using a medical text pre-training model and a multi-layer drug knowledge graph, and is represented and learned in combination with a heterogeneous graph neural network, which can more effectively extract data features and improve the accuracy and efficiency of signal mining. The application of the multi-task learning framework can simultaneously predict the drug-adverse reaction association probability and risk level, further improving the prediction performance of the model. The dynamic change characteristics of the association probability and risk level are detected by the time series analysis method, and the normalized drug adverse reaction data are combined with the early warning analysis model for real-time risk early warning, which can realize the dynamic monitoring and early warning of drug safety signals and timely discover potential risks. The hierarchical evaluation system and the setting of differentiated monitoring cycles and trigger thresholds can be refined according to the risk level, improving the pertinence and effectiveness of early warning. The construction of a multi-layer drug knowledge graph integrates drug molecular feature information and clinical medication information, which can provide richer contextual information and enhance the interpretability of the signal. The candidate signal can be verified in combination with clinical data to further improve the reliability of the signal. The application of the hierarchical evaluation system also makes the evaluation results of the signal clearer and more understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A schematic diagram of the process of a method for data collection and drug safety signal mining for PMS according to an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of the structure of the data collection and drug safety signal mining agent for PMS according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0074] Figure 1 FIG. 1 is a flow chart of a method for data collection and drug safety signal mining for PMS according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0075] S101. Acquire historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping on drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by integrating drug molecular feature information and clinical medication information. The multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes;

[0076] S102. A heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, a node type-aware sampling strategy is used to construct training samples, and drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network to obtain a node vector representation of multidimensional information fusion. Based on the node vector representation, a multi-task learning framework is constructed. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identifier;

[0077] S103. Construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized drug adverse reaction data.

[0078] The present invention extracts historical monitoring data from the PMS system through a data acquisition unit, and the historical monitoring data includes drug use records and drug adverse reaction records. The drug use records include but are not limited to information such as the generic name of the drug, medication time, medication population, and medication dosage; the drug adverse reaction records include but are not limited to information such as adverse reaction description, occurrence time, and severity.

[0079] To ensure data quality, the present invention uses a medical text pre-training model pre-trained on a medical literature corpus to perform data standardization. Specifically:

[0080] For drug names, different forms of drug names such as trade names and chemical names are uniformly mapped to standardized generic drug names. This mapping process is completed based on the name correspondence established in the drug instructions database to ensure that the same drug from different sources can be accurately identified.

[0081] For adverse reaction descriptions, the adverse reaction symptoms described in natural language are mapped to standardized medical vocabulary terms. This mapping uses the preferred term (PT) in the medical terminology standard vocabulary MedDRA (Medical Dictionary for Regulatory Activities) as the standard terminology, and the non-standard symptom descriptions are automatically converted into standard terms through the pre-trained model.

[0082] After the above standardization process, a standardized drug adverse reaction data set is formed. Each record in this data set contains standardized drug names and adverse reaction terms, providing a standardized data basis for the subsequent construction of the knowledge graph. Through this standardization process, problems such as inconsistent names and non-standard descriptions in the original data can be effectively solved, and data quality and availability can be improved.

[0083] In an optional embodiment, the drug name and adverse reaction description in the original data are standardized and mapped using a medical text pre-training model to generate standardized adverse drug reaction data, and based on the standardized adverse drug reaction data, a multi-layer drug knowledge graph is constructed by integrating drug molecular feature information and clinical medication information, including:

[0084] Inputting the original adverse drug reaction data into a medical text pre-training model, the medical text pre-training model performs word segmentation on the input text and extracts word vectors and term vectors, concatenating the word vectors with the term vectors to form a feature vector, inputting the feature vector into a bidirectional long short-term memory unit to extract context features, and performing sequence labeling through a conditional random field based on the context features to obtain initial drug entities and initial adverse reaction entities;

[0085] Calculating the string similarity and semantic similarity between the initial drug entity and the standard medical term, determining the initial drug entity whose weighted combination value of the two similarities exceeds the mapping threshold as the standard drug entity, and mapping the initial adverse reaction entity to the standard adverse reaction entity to generate standardized drug adverse reaction data;

[0086] The simplified linear input canonical representation of the drug molecule is converted into a molecular structure diagram with atom type and chemical bond type annotations. The molecular structure diagram is subjected to local feature extraction through multi-layer spatial convolution, and the extracted local features are aggregated through attention weights to obtain a drug molecule feature vector;

[0087] Input clinical medication text into the bidirectional attention network, calculate the attention weight with standard medical terms, and obtain standardized clinical medication entities;

[0088] The sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; the molecular similarity is calculated using drug molecular feature vectors to construct a molecular structure association network; and a clinical association network is constructed based on standardized clinical drug entities;

[0089] Input the temporal association network, the molecular structure association network and the clinical association network into a hierarchical graph attention network, generate a node vector representation through a random walk algorithm, the node vector representation obtains multi-layer network features through a graph attention layer, and the multi-layer network features are input into a recurrent neural network to extract temporal features;

[0090] The multi-layer network features and time series features are input into a heterogeneous graph neural network, and a multi-layer drug knowledge graph containing node vector representations is obtained through feature transformation training to form a multi-dimensional association network of drug information nodes and adverse reaction information nodes.

[0091] A method for constructing a multi-layer drug knowledge graph using a medical text pre-training model, drug molecular feature information, and clinical medication information for drug adverse reaction association analysis is implemented as follows:

[0092] First, the original adverse drug reaction data is normalized. The original data is input into a pre-trained medical text model (such as BERT, BioBERT). The model segmented the input text and extracted word vectors and word vectors respectively. The word vectors and word vectors are concatenated to form a feature vector. The feature vector is input into a bidirectional long short-term memory unit (BiLSTM) network to extract contextual features. Then, the conditional random field (CRF) model is used for sequence labeling to identify the initial drug entity and the initial adverse reaction entity. For example, for the input text "Headache after taking aspirin", the model may identify "aspirin" as the initial drug entity and "headache" as the initial adverse reaction entity.

[0093] Next, the identified initial drug entities and adverse reaction entities are standardized and mapped. By calculating the string similarity (e.g., edit distance, Jaro-Winkler similarity) and semantic similarity (e.g., cosine similarity calculated based on Word2Vec or a pre-trained model) between the initial drug entity and the drug name in the standard medical terminology library (e.g., DrugBank, SNOMED CT), the initial drug entity whose weighted combination value of the two similarities exceeds a preset threshold (e.g., 0.9) is determined as a standard drug entity. At the same time, the initial adverse reaction entity is mapped to the adverse reaction term in the standard medical terminology library to form a standard adverse reaction entity. For example, "aspirin" is mapped to "Aspirin" and "headache" is mapped to "Headache". Finally, standardized drug adverse reaction data is generated.

[0094] Then, extract the molecular features of the drug. Convert the molecular simplified linear input specification (SMILES) string of the drug into a molecular structure diagram with atom type and chemical bond type annotations. The molecular structure diagram is input into a multi-layer spatial convolutional neural network (such as Graph Convolutional Network, GCN) for local feature extraction. The extracted local features are weighted and aggregated through the attention mechanism to obtain the drug molecular feature vector. For example, after the molecular structure diagram of aspirin passes through GCN and the attention mechanism, a fixed-dimensional vector is obtained to represent its molecular features.

[0095] At the same time, the clinical medication information is standardized. The clinical medication text is input into the bidirectional attention network, and the attention weight of each word in the text and the term in the standard medical terminology library is calculated. According to the attention weight, the standardized clinical medication entities are extracted. For example, the input text "patients take 100 mg of aspirin orally every day" can extract standardized clinical medication entities such as "aspirin" and "oral".

[0096] Next, a multi-layer drug knowledge graph is constructed. A sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities in the standardized drug adverse reaction data, and a temporal association network is constructed. Molecular similarity (such as cosine similarity, Tanimoto coefficient) is calculated using drug molecular feature vectors to construct a molecular structure association network. A clinical association network is constructed based on standardized clinical drug entities. These three networks are input into a hierarchical graph attention network. A node vector representation is generated through a random walk algorithm. The node vector representation obtains multi-layer network features through the graph attention layer. Multi-layer network features are input into a recurrent neural network (such as GRU) to extract temporal features.

[0097] Finally, the multi-layer network features and time series features are input into a heterogeneous graph neural network (such as Heterogeneous Graph Attention Network, HAN), and a multi-layer drug knowledge graph containing node vector representation is obtained through feature transformation training, forming a multi-dimensional association network of drug information nodes and adverse reaction information nodes. For example, there is an edge between the aspirin node and the headache node, indicating that there is an association between the two, and the weight of the edge represents the strength of the association.

[0098] In this embodiment, by combining the medical text pre-training model, sequence annotation technology and standardized mapping method, it is possible to more accurately identify and normalize adverse drug reaction information, reducing the workload of manual review. By integrating drug molecular characteristics, clinical medication information and time-series association information, a multi-layer drug knowledge graph is constructed, which can analyze the relationship between drugs and adverse reactions from multiple perspectives and provide more comprehensive information support. By using heterogeneous graph neural networks and graph attention mechanisms, it is possible to effectively learn complex relationships in drug knowledge graphs, improve the efficiency and accuracy of adverse drug reaction prediction, and provide a powerful tool for drug safety assessment.

[0099] In an optional embodiment, a sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; a drug molecule feature vector is used to calculate the molecular similarity to construct a molecular structure association network; and a clinical association network is constructed according to standardized clinical drug entities, including:

[0100] A sliding time window is used to detect the co-occurrence relationship between the standard drug entity and the standard adverse reaction entity in the time dimension to form a drug-adverse reaction entity pair, and a temporal co-occurrence metric value of the drug-adverse reaction entity pair is calculated for each time window. The temporal co-occurrence metric value is determined by the ratio of the number of co-occurrences of the drug-adverse reaction entity pair in the window to the total number of occurrences of each drug-adverse reaction entity pair;

[0101] Performing a temporal significance test on the drug-adverse reaction entity pairs, generating a co-occurrence probability distribution by a random permutation method, establishing connections between the drug-adverse reaction entity pairs that pass the significance test, and using the temporal co-occurrence measurement value as a connection weight value to construct a temporal association network;

[0102] Inputting the drug molecule feature vector into a similarity calculation model, the similarity calculation model performs attention weighted calculation on the chemical substructures in the drug molecule feature vector, outputs a molecular similarity matrix, inputs the molecular similarity matrix into a spectral clustering model for drug grouping, and connects drug molecules with similarities higher than a preset threshold according to the grouping results to construct a molecular structure association network;

[0103] Forming a drug combination sequence according to the clinical drug entity, calculating the co-occurrence times of drugs in the drug combination sequence, obtaining the clinical co-occurrence support of the drug combination, establishing a Markov chain model for the drug combination sequence in time sequence, and calculating the transition probability between drugs in adjacent time sequence based on the Markov chain model;

[0104] Constructing a bipartite graph of drugs and clinical indications, calculating the drug-clinical indication co-occurrence probability, calculating the mutual information ratio of the drug edge probability and the clinical indication edge probability according to the drug-clinical indication co-occurrence probability to obtain the drug-clinical indication association strength, and weightedly combining the clinical co-occurrence support, drug transfer probability and drug-clinical indication association strength as the connection weight value to construct a clinical association network;

[0105] The connection weights in the temporal association network, the connection weights in the molecular structure association network, and the connection weights in the clinical association network are fused to construct a multi-layer drug association network.

[0106] Exemplarily, first, standardized drug entities and standardized adverse reaction entities are extracted from a large amount of electronic medical record data. A sliding time window, for example, set to 30 days, is used to detect the co-occurrence relationship of these entities in the time dimension. For example, if a certain drug and a certain adverse reaction appear multiple times in the same 30-day window, they are considered to have a temporal co-occurrence relationship. In order to quantify this relationship, the temporal co-occurrence metric of the drug-adverse reaction entity pair in each time window is calculated. This metric can be determined by dividing the number of co-occurrences of the drug-adverse reaction entity pair in the window by the total number of occurrences of each. For example, if drug A and adverse reaction B appear together 5 times in 30 days, drug A appears a total of 20 times, and adverse reaction B appears a total of 10 times, then their temporal co-occurrence metric is 5 / (20+10)=0.167.

[0107] Next, in order to determine whether these co-occurrence relationships are statistically significant, a temporal significance test is required. A random permutation method is used to generate a co-occurrence probability distribution to test whether the observed co-occurrence relationship is significantly higher than the random expectation. Specifically, the drug and adverse reaction entities in the original data are randomly shuffled to simulate random co-occurrence situations, and repeated multiple times, such as 1,000 times, to obtain a co-occurrence probability distribution. If the actual observed co-occurrence metric is greater than 95% of the random permutation results, the co-occurrence relationship is considered to be statistically significant. The drug-adverse reaction entity pairs that pass the significance test are connected, and the calculated temporal co-occurrence metric is used as the weight value of the connection to construct a temporal association network.

[0108] At the same time, the molecular structure information of the drug is used to construct a molecular structure association network. First, the drug molecules are converted into feature vectors, for example, using the extended connectivity fingerprint ECFP4. These feature vectors are input into a similarity calculation model. The model performs attention weighting calculations on the chemical substructures in the feature vectors, for example, focusing on those substructures that are critical to the activity of the drug. The model outputs a molecular similarity matrix that reflects the structural similarities between different drug molecules. This similarity matrix is ​​then input into a spectral clustering model to group the drug molecules. Based on the grouping results, drug molecules with a similarity higher than a preset threshold (for example, 0.8) are connected to construct a molecular structure association network.

[0109] In addition, clinical medication information is used to construct a clinical association network. Based on the standardized clinical medication entities, for example, the names of drugs extracted from the hospital's prescription data, a drug combination sequence is formed. The number of co-occurrences of drugs in the drug combination sequence is calculated to obtain the clinical co-occurrence support of the drug combination. For example, if drug A and drug B often appear together in the prescription of the same patient, their clinical co-occurrence support is high. The drug combination sequence is built into a Markov chain model in chronological order, and the transition probability between adjacent time series drugs is calculated based on the model. For example, the probability of using drug B after using drug A is calculated. In addition, a bipartite graph of drugs and clinical indications is constructed to represent the correspondence between drugs and clinical indications, and the drug-clinical indication co-occurrence probability is calculated. For example, count how many patients have used a certain drug at the same time and have been diagnosed with a certain disease. According to the co-occurrence probability, the mutual information ratio of the drug marginal probability and the clinical indication marginal probability is calculated to obtain the drug-clinical indication association strength. Finally, the clinical co-occurrence support, drug transition probability, and drug-clinical indication association strength are weighted and combined as the connection weight value to construct a clinical association network.

[0110] Finally, the connection weights in the temporal association network, the connection weights in the molecular structure association network, and the connection weights in the clinical association network are fused to construct a multi-layer drug association network. For example, the weight values ​​of the three networks can be linearly combined, or a more complex fusion strategy can be used.

[0111] In this embodiment, by using a multi-layer drug knowledge graph to fuse drug molecular features and clinical medication information, and combining heterogeneous graph neural networks for representation learning, it is possible to fully utilize the multi-dimensional information of drugs and improve the accuracy of safety signal detection. Based on time series feature extraction and multi-task learning framework, the drug-adverse reaction association probability and risk level can be predicted simultaneously, effectively capturing the trend of risk changes. The establishment of a graded evaluation system realizes the systematic evaluation of safety signals, and differentiated monitoring strategies improve regulatory efficiency. The dynamically decreasing exploration probability strategy ensures that the model has sufficient exploration space in the early stages of training, and gradually converges to the optimal strategy as training proceeds. This solution significantly improves the timeliness and accuracy of drug safety signal warnings, reduces the false alarm rate, and achieves optimal allocation of regulatory resources through standardized processing of raw data and fusion learning of multi-dimensional features.

[0112] In an optional embodiment, a heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, a node type-aware sampling strategy is used to construct training samples, and drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network. The node vector representation of multidimensional information fusion is obtained, which includes:

[0113] Receiving a multidimensional association network and a drug action mechanism rule, performing a linear transformation on a node feature vector in the multidimensional association network, and generating a query vector, a key vector, and a value vector;

[0114] An attention score matrix is ​​obtained by performing a dot product operation on the query vector and the key vector, the attention score matrix is ​​normalized to obtain a temporal attention weight, a matrix multiplication operation is performed on the temporal attention weight and the value vector to obtain a temporal context vector, and the temporal context vector is concatenated with the node feature vector and then the temporal enhancement feature is obtained through a feedforward neural network;

[0115] The time series enhancement feature is weightedly summed with the edge weight in the multidimensional association network to obtain the edge weight of the fused time series feature, the degree centrality of each node is calculated as the node importance score according to the edge weight of the fused time series feature, the nodes are divided into different categories according to the node type, and probability sampling is performed in each category according to the node importance score to obtain a training node set;

[0116] Extracting drug nodes associated with the mechanism of action from the drug action mechanism rules as drug node pairs, calculating the cosine similarity of the drug node pairs in the training node set, and taking the difference between the similarity and a preset node similarity threshold as the rule loss;

[0117] Input the training node set into the heterogeneous graph neural network for node classification, use cross entropy to calculate the classification loss, and use the weighted sum of the classification loss and the rule loss as the total loss function;

[0118] Based on the total loss function, the parameters of the heterogeneous graph neural network are gradient updated, and the iteration is repeated until the preset number of iterations is met to obtain the final node vector representation of multi-dimensional information fusion.

[0119] Exemplarily, first, prepare multidimensional association network data and drug action mechanism rules. Multidimensional association network data contains different types of nodes (e.g., drugs, genes, diseases) and multiple association relationships between them (e.g., drug-target interaction, gene-disease association). Each node has an initial feature vector, such as the chemical structure characteristics of the drug, the expression spectrum characteristics of the gene, etc. The drug action mechanism rules describe the association between drugs due to the common action mechanism. For example, if two drugs act on the same target or pathway, they are considered to have an action mechanism association.

[0120] Next, the node feature vector is linearly transformed to generate the query vector, key vector, and value vector for attention mechanism calculation. For example, the node feature vector can be transformed using three different linear transformation matrices to obtain the corresponding query vector, key vector, and value vector.

[0121] Then, the attention score matrix is ​​obtained by performing a dot product operation on the query vector and the key vector. For example, the query vector of each node is dot-producted with the key vectors of all other nodes to obtain a matrix where each element represents the attention score between two nodes.

[0122] The attention score matrix is ​​normalized to obtain the temporal attention weight. For example, a softmax function can be used to normalize each row of the attention score matrix so that the sum of the elements in each row is 1.

[0123] The temporal attention weight is matrix multiplied with the value vector to obtain the temporal context vector. The temporal context vector integrates the information of the node neighbors and takes into account the importance of different neighbor nodes.

[0124] After concatenating the temporal context vector with the node feature vector, the temporal enhancement feature is obtained through a feedforward neural network. The feedforward neural network can be a multi-layer perceptron, which is used to further extract and fuse the feature information of the node.

[0125] The time series enhancement feature and the edge weight in the multidimensional association network are weighted and summed to obtain the edge weight of the fused time series feature. For example, the time series enhancement feature can be multiplied by the original edge weight to obtain a new edge weight.

[0126] The degree centrality of each node is calculated as the node importance score according to the edge weight of the fused time series features. Degree centrality indicates the importance of a node in the network.

[0127] Nodes are divided into different categories according to their types. For example, drugs, genes, and diseases can be divided into different categories.

[0128] In each category, probability sampling is performed according to the node importance score to obtain a set of training nodes. For example, a weighted random sampling method can be used to select nodes with higher importance scores. Suppose there are three drug nodes A, B, and C, with importance scores of 0.8, 0.5, and 0.2, respectively. If two nodes are to be sampled, the probability of node A being selected is 0.8 / (0.8+0.5+0.2)=0.5, the probability of node B being selected is 0.31, and the probability of node C being selected is 0.19.

[0129] Drug nodes with associated action mechanisms are extracted from the drug action mechanism rules as drug node pairs. For example, if drug A and drug B act on the same target, (A, B) is taken as a drug node pair.

[0130] Calculate the cosine similarity of drug node pairs in the training node set. Cosine similarity is used to measure the similarity between two node vectors.

[0131] The difference between the similarity and the preset node similarity threshold is used as the rule loss. For example, if the node similarity is lower than the threshold, the rule loss is positive, encouraging the model to learn a higher similarity. Assuming the preset similarity threshold is 0.7 and the similarity of the drug node pair (A, B) is 0.6, the rule loss is 0.7-0.6=0.1.

[0132] The training node set is input into the heterogeneous graph neural network for node classification. The heterogeneous graph neural network can learn the complex relationships between different types of nodes.

[0133] The classification loss is calculated using cross entropy, which is used to measure the difference between the predicted result and the true label.

[0134] The weighted sum of the classification loss and the rule loss is used as the total loss function. For example, the two loss functions can be multiplied by different weights and then added together.

[0135] Based on the total loss function, the parameters of the heterogeneous graph neural network are gradient updated, and the iteration is repeated until the preset number of iterations is met to obtain the final node vector representation of multi-dimensional information fusion.

[0136] In this embodiment, by fusing time series information and drug action mechanism rules, the method can more accurately learn the vector representation of nodes, thereby improving the prediction performance of downstream tasks (such as drug target prediction, drug repositioning). The attention mechanism is used to learn the association between nodes, and combined with the drug action mechanism rules, the prediction results of the model are more interpretable, which helps to understand the drug's mechanism of action. The node type-aware sampling strategy is adopted to reduce the number of training samples and improve the training efficiency of the model.

[0137] In an optional embodiment, based on the node vector representation, a multi-task learning framework is constructed, the multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification, including:

[0138] Obtaining and concatenating the drug node vector and the adverse reaction node vector in the node vector representation, extracting features through a multi-layer attention-enhanced feedforward neural network with residual connections to generate a drug-adverse reaction shared feature vector;

[0139] Constructing a time series feature sequence matrix, wherein the rows of the time series feature sequence matrix represent different time windows, and the columns represent feature sets corresponding to the drug node vector and the adverse reaction node vector in each time window, wherein the feature set includes statistical association features calculated based on an adaptive time window, drug molecular structure features, and drug target features;

[0140] Inputting the temporal feature sequence matrix into a bidirectional long short-term memory network provided with a gating mechanism, wherein the gating mechanism adaptively adjusts the feature weight according to the importance of the time window to generate a temporal context vector;

[0141] A multi-head attention mechanism is used to fuse the drug-adverse reaction shared feature vector with the temporal context vector to obtain a drug-adverse reaction multi-task feature vector, and the drug-adverse reaction multi-task feature vector is input into an association prediction network and a risk prediction network respectively to generate an initial drug-adverse reaction association probability and an initial risk level probability;

[0142] Calculate the binary cross entropy loss based on the initial drug-adverse reaction association probability, calculate the multi-category cross entropy loss based on the initial risk level probability, construct a multi-task loss function in combination with the time series smoothing constraint, update the model parameters by iteratively optimizing the multi-task loss function, and obtain the optimized drug-adverse reaction association probability and risk level probability;

[0143] The kernel density estimation method is used to calculate the probability distribution of the current time window based on the optimized drug-adverse reaction association probability, determine the dynamic association threshold, and screen candidate drug safety signals from the optimized drug-adverse reaction association probability and risk level probability according to the dynamic association threshold and the preset risk threshold.

[0144] Exemplarily, first, a knowledge graph containing drug and adverse reaction information is constructed. For example, information from public databases (such as DrugBank, SIDER, MedDRA) can be integrated, where nodes represent drugs or adverse reactions and edges represent known relationships between them. Then, using a graph embedding algorithm such as Node2Vec or TransE, each node in the graph is converted into a low-dimensional vector representation to capture its semantic information. Assume that the vector representation of drug A is [0.1, 0.2, 0.3] and the vector representation of adverse reaction B is [0.4, 0.5, 0.6].

[0145] Concatenate the node vectors of drugs and adverse reactions and perform feature extraction. Concatenate the obtained drug node vectors and adverse reaction node vectors together. For example, concatenate the vector [0.1, 0.2, 0.3] of drug A and the vector [0.4, 0.5, 0.6] of adverse reaction B into [0.1, 0.2, 0.3, 0.4, 0.5, 0.6]. Then, input the concatenated vector into a multi-layer attention-enhanced feedforward neural network. The network contains multiple fully connected layers and residual connections, and an attention mechanism is added to each layer to highlight important features. Assume that after passing through the network, the output shared feature vector is [0.7, 0.8, 0.9].

[0146] Construct a time series feature sequence matrix. Collect drug and adverse reaction data in different time windows. For example, collect data for the past 24 months in units of months. For each time window, calculate the statistical association features between drugs and adverse reactions, such as the number of reports, reporting ratios, etc. At the same time, extract the molecular structure features of the drug (such as molecular fingerprints) and the drug target features. Combine these features into a feature set. Suppose that in a certain time window, the feature set of drug A and adverse reaction B is [100, 0.01, 1, 0], where 100 represents the number of co-occurrences, 0.01 represents the reporting ratio, 1 represents the presence of a molecular structure feature, and 0 represents the absence of a target feature. Arrange the feature sets of different time windows into a matrix, where each row represents a time window and each column represents a feature.

[0147] Generate a time series context vector. Input the constructed time series feature sequence matrix into a bidirectional long short-term memory network (Bi-LSTM) with a gating mechanism. The gating mechanism adaptively adjusts the feature weights according to the importance of the time window. For example, recent data may be more important than distant data. The Bi-LSTM network can capture long-term dependencies in time series data. Assume that after passing through the Bi-LSTM network, the generated time series context vector is [0.1, 0.2].

[0148] Fusion of shared feature vectors and temporal context vectors. Use multi-head attention mechanism to fuse drug-adverse reaction shared feature vectors and temporal context vectors. Multi-head attention mechanism can capture the relationship between two vectors from different perspectives. Assume that the fused drug-adverse reaction multi-task feature vector is [0.3, 0.4, 0.5].

[0149] Predict association probability and risk level. The fused multi-task feature vector is input into the association prediction network and the risk prediction network respectively. Both networks are fully connected neural networks. The association prediction network outputs the association probability of drug-adverse reaction, and the risk prediction network outputs the probability distribution of risk level. Assume that the association probability between drug A and adverse reaction B is 0.9, and the probability distribution of risk level is [0.1, 0.8, 0.1], which indicates the probability of low, medium and high risk levels.

[0150] Optimize model parameters. Based on the predicted association probability and risk level probability, calculate the binary cross entropy loss and multi-category cross entropy loss. At the same time, add a time series smoothing constraint to make the prediction results of adjacent time windows change smoothly. Combine these losses into a multi-task loss function and use an optimization algorithm (such as Adam optimizer) to iteratively update the model parameters until the loss function converges.

[0151] Generate candidate drug safety signals. Using the kernel density estimation method, calculate the probability distribution of the current time window based on the optimized drug-adverse reaction association probability, and determine the dynamic association threshold. For example, set the threshold to the 95% quantile of the probability distribution. According to the dynamic association threshold and the preset risk threshold (e.g., medium risk and above), screen candidate drug safety signals from the optimized association probability and risk level probability. For example, if the association probability between drug A and adverse reaction B is greater than the dynamic association threshold, and the risk level is medium or high, it is marked as a candidate drug safety signal.

[0152] In this embodiment, by combining multi-task learning, time series analysis and knowledge graph, the relationship between drugs and adverse reactions can be considered more comprehensively, thereby improving the accuracy of prediction. Through time series analysis, the dynamic changes of drug-adverse reaction association probability and risk level can be captured, thereby achieving real-time monitoring of drug safety. The model can not only predict the drug-adverse reaction association probability, but also predict the risk level, providing more refined risk information for regulators and medical professionals.

[0153] In an optional embodiment, a hierarchical evaluation system including a signal strength dimension and a clinical relevance dimension is constructed, the candidate drug safety signal is verified based on the hierarchical evaluation system, the reliability of the candidate drug safety signal is confirmed in combination with clinical data, a target drug safety signal is output, the target drug safety signal is mapped to the hierarchical evaluation system, an early warning level is determined based on the mapping position, a differentiated monitoring period and a trigger threshold are set according to the early warning level, and the node vector representation is input into an early warning analysis model, and real-time risk early warning of the standardized adverse drug reaction data includes:

[0154] Construct a hierarchical evaluation system of signal strength dimension and clinical relevance dimension. The signal strength dimension is based on the strength of drug-adverse reaction causal relationship, the number of adverse reaction reports, and the time trend of adverse reactions to construct a weighted scoring model to obtain a signal strength score. The clinical relevance dimension is based on the severity of adverse reactions, population sensitivity, and clinical interventionability to construct a hierarchical analysis network to obtain a clinical relevance score.

[0155] Inputting the candidate drug safety signal, signal strength score, and clinical relevance score into the screening model to obtain a first screening result, inputting the first screening result into a cohort analysis unit, a literature analysis unit, and a social media analysis unit, respectively, and combining the verification results of each unit using an integrated learning model to obtain a multi-source verification result;

[0156] Inputting the drug molecular characteristics, metabolic information, action network information and the multi-source verification results corresponding to the candidate drug safety signal into a graph neural network model, and determining the target drug safety signal based on the output results of the graph neural network model;

[0157] Inputting the target drug safety signal, the signal strength score, and the clinical relevance score into a pre-trained deep reinforcement learning model to obtain a warning level;

[0158] A monitoring strategy is constructed according to the warning level, and the monitoring strategy and the time series characteristics of the target drug safety signal are input into a hybrid neural network model. The time series characteristics include the number characteristics of adverse reaction reports, the cumulative proportion characteristics and the geographical distribution characteristics. Based on the output results of the hybrid neural network model, risk warnings are issued for adverse drug reaction data.

[0159] Exemplarily, first, a hierarchical evaluation system is constructed. This system contains two dimensions: signal strength dimension and clinical relevance dimension. For the signal strength dimension, three factors are considered: the strength of the causal relationship between the drug and the adverse reaction, the number of adverse reaction reports, and the trend of adverse reactions over time. Based on these three factors, a weighted scoring model is established to calculate the signal strength score. For example, the strength of the causal relationship can be divided into four levels: very strong, strong, medium, and weak, and 4, 3, 2, and 1 points are assigned respectively; the number of reports can be set with different scores according to the magnitude difference; the time trend can be set with different scores according to the rising speed. The three scores are weighted and summed to obtain the final signal strength score. For the clinical relevance dimension, the severity of adverse reactions, the sensitivity of the affected population, and the feasibility of clinical intervention are considered. The network is constructed using the hierarchical analysis method to determine the weights of each factor, and finally the clinical relevance score is calculated. For example, the severity can be divided into four levels: fatal, severe, moderate, and mild, and 4, 3, 2, and 1 points are assigned respectively; the sensitivity of the population can be assigned different scores according to whether special populations (such as children and the elderly) are affected; clinical interventionability can be assigned different scores according to whether there are effective treatments. These scores were weighted and summed to obtain the final clinical relevance score.

[0160] Next, the candidate drug safety signals are verified. The candidate signals, signal strength scores, and clinical relevance scores are input into the screening model to obtain preliminary screening results. Then, the preliminary screening results are input into the cohort analysis unit, literature analysis unit, and social media analysis unit for further verification. The cohort analysis unit analyzes the relationship between drug exposure and the incidence of adverse reactions in a specific population based on real-world data; the literature analysis unit retrieves published medical literature to find evidence related to the candidate signals; and the social media analysis unit analyzes discussions on social media platforms to identify potential safety risks. An integrated learning model, such as a random forest or a gradient boosting decision tree, is used to integrate the verification results of the three units to obtain multi-source verification results.

[0161] Then, the target drug safety signal is determined. The molecular characteristics (such as chemical structure, target information), metabolic information (such as absorption, distribution, metabolism, excretion) and action network information (such as interaction with other drugs or biological molecules) of the candidate drugs are input into the graph neural network model together with the multi-source verification results. Graph neural networks can effectively process complex network structure data and learn the patterns and relationships therein. Based on the output results of the graph neural network model, the final target drug safety signal is determined.

[0162] After the target signal is determined, the warning level is evaluated. The target signal, signal strength score, and clinical relevance score are input into a pre-trained deep reinforcement learning model. The model learns the best strategy through interaction with the environment, thereby determining the most appropriate warning level based on the input features. The warning level can be divided into different levels, such as level 1, level 2, and level 3, corresponding to different risk levels.

[0163] Finally, set up monitoring strategies and conduct real-time risk warnings based on the warning levels. Set different monitoring cycles and trigger thresholds based on different warning levels. For example, high-level warnings may require shorter monitoring cycles and lower trigger thresholds. Input the monitoring strategy and the timing characteristics of the target signal (such as the number of adverse reaction reports, cumulative proportions, and geographical distribution) into the hybrid neural network model. The hybrid neural network model can combine different types of data, such as time series data and static data, to make more accurate predictions. Based on the output results of the hybrid neural network model, conduct real-time risk warnings for adverse drug reaction data. For example, if the model predicts that the number of adverse reaction reports will exceed a preset threshold in the future, an early warning is issued.

[0164] Assume that the number of adverse reaction reports of a certain drug with rash increases. The signal strength score calculated by the above method is 8, and the clinical relevance score is 6. After multi-source verification, it is confirmed as a target drug safety signal. The deep reinforcement learning model evaluates it as a secondary warning, sets the monitoring cycle to once a month, and the trigger threshold is a 50% increase in the number of reports. Subsequent monitoring found that the number of adverse reaction reports of the drug rash increased by 60% in one month, exceeding the preset threshold. The system issued an early warning, prompting relevant departments to take further investigation and intervention measures.

[0165] In this embodiment, through multi-dimensional evaluation and multi-source data verification, it is possible to more accurately identify true safety signals, reduce false positives and missed positives, and improve the efficiency of signal detection. Using deep learning and reinforcement learning techniques, it is possible to monitor and predict adverse drug reaction data in real time, detect potential risks in a timely manner, and issue warnings to provide support for the decision-making of relevant departments. Differentiated monitoring cycles and trigger thresholds are set according to different warning levels to achieve personalized monitoring strategies, improve resource utilization efficiency, and avoid the burden of excessive monitoring.

[0166] In an optional embodiment, the target drug safety signal, the signal strength score and the clinical relevance score are input into a pre-trained deep reinforcement learning model to obtain an early warning level including:

[0167] Collect characteristic information of the target drug safety signal, the characteristic information including signal structured representation, signal intensity temporal variation characteristics, clinical relevance score characteristics and historical warning records, construct the characteristic information into a state space and map it into a state vector;

[0168] Constructing an action space for warning level adjustment, the action space including a warning level raising action, a warning level lowering action and a warning level maintaining action, mapping the action space into an action vector, and the action vector and the state vector form a state-action pair;

[0169] Constructing a reward function for early warning evaluation, calculating a weighted reward value based on an early warning accuracy index, an early warning timeliness index, and a supervision resource consumption index, and forming a training sample with the state-action pair and the weighted reward value;

[0170] Constructing a dual learning network structure including a current network and a target network, storing the training samples in an experience replay pool, sampling from the experience replay pool to obtain a state transition sequence, inputting the state transition sequence into the dual learning network structure, and using a temporal difference algorithm to alternately optimize and update the parameters of the current network and the target network;

[0171] A decreasing greedy strategy is used for action selection. The corresponding action is selected based on the current state and environmental feedback is obtained. The environmental feedback is input into the dual learning network structure to obtain the optimal early warning decision strategy. The real-time feature vector of the target drug safety signal is input into the optimal early warning decision strategy, and the corresponding early warning level is output.

[0172] A drug safety early warning method and system are used to evaluate the safety signals of target drugs and give corresponding early warning levels.

[0173] First, the characteristic information of the target drug safety signal is collected. These characteristic information covers multiple dimensions, including: structured representation of the signal, for example, describing the adverse event report in a structured form, including patient information, medication information, adverse event information, etc.; signal strength time series change characteristics, for example, statistically analyzing the number of adverse event reports over a period of time, for example, using the moving average method to smooth the number of reports over a period of time to reduce the impact of random fluctuations on trend judgment; clinical relevance score characteristics, for example, experts conduct a comprehensive assessment based on the severity, incidence, and relevance of adverse events, and give a score of 0-10 points. The higher the score, the stronger the clinical relevance; and historical warning records, for example, recording the changes in the warning level of the drug over the past period of time. Taking a certain cold medicine as an example, assuming that the number of adverse event reports of the drug continues to rise over a period of time, clinical experts evaluate its relevance score as 8 points, and the drug has not issued a warning recently, then this information will constitute the characteristic information of the drug safety signal.

[0174] Next, the collected feature information is constructed into a state space and mapped into a state vector. For example, the signal structured representation, signal strength temporal variation characteristics, clinical relevance scoring characteristics, and historical warning records can be quantized into numerical values, and then these numerical values ​​are combined into a vector, which is the state vector. For example, assuming that the signal structured representation is quantized to 0.8, the signal strength temporal variation characteristics are quantized to 0.9, the clinical relevance scoring characteristics are quantized to 8, and the historical warning records are quantized to 0, the state vector can be expressed as [0.8, 0.9, 8, 0].

[0175] Then, construct an action space for warning level adjustment. The action space includes warning level increase actions, such as increasing the warning level from "blue" to "yellow"; warning level decrease actions, such as decreasing the warning level from "yellow" to "blue"; and warning level maintenance actions, such as maintaining the current "green" warning level. Map these actions to action vectors. For example, [1, 0, 0] can be used to represent the warning level increase action, [0, 1, 0] can be used to represent the warning level decrease action, and [0, 0, 1] can be used to represent the warning level maintenance action. The state vector and the action vector form a state-action pair. For example, [0.8, 0.9, 8, 0] and [1, 0, 0] form a state-action pair, which means that the warning level increase action is performed in the current state.

[0176] Next, construct a reward function for early warning evaluation. The reward function is based on the early warning accuracy index, for example, evaluating the degree of match between the early warning level and the actual drug safety risk; the early warning timeliness index, for example, evaluating the interval between the time the early warning is issued and the time when the adverse event occurs; and the regulatory resource consumption index, for example, evaluating the cost required for early warning processing, to calculate the weighted reward value. The state-action pair and the weighted reward value constitute a training sample. For example, assuming that the early warning accuracy index of a state-action pair is 0.9, the early warning timeliness index is 0.8, and the regulatory resource consumption index is 0.7, the weighted reward value can be calculated according to the preset weights, for example, the weighted reward value = 0.5*0.9+0.3*0.8+0.2*0.7 = 0.83. Then ([0.8, 0.9, 8, 0], [1, 0, 0], 0.83) constitutes a training sample.

[0177] Then, a dual learning network structure including the current network and the target network is constructed. The training samples are stored in the experience replay pool. The state transition sequence is sampled from the experience replay pool. The state transition sequence is input into the dual learning network structure, and the parameters of the current network and the target network are alternately optimized and updated using the temporal difference algorithm.

[0178] Use a decreasing greedy strategy for action selection. Select the corresponding action based on the current state and obtain environmental feedback. Input the environmental feedback into the dual learning network structure to obtain the optimal early warning decision strategy. Input the real-time feature vector of the target drug safety signal into the optimal early warning decision strategy and output the corresponding warning level. For example, input the real-time feature vector [0.7, 0.8, 7, 1] into the optimal early warning decision strategy and output the warning level "yellow".

[0179] In this embodiment, through the training and optimization of the deep reinforcement learning model, the drug safety risk can be more accurately assessed, the accuracy of the warning level can be improved, and the safety of public drug use can be better guaranteed. Using real-time feature vectors for early warning decisions can more quickly identify potential drug safety risks, issue early warnings in a timely manner, and avoid further expansion of adverse events. By considering regulatory resource consumption indicators, the configuration of early warning resources can be optimized, resource waste can be avoided, and regulatory efficiency can be improved.

[0180] Figure 2 FIG. 1 is a schematic diagram of the structure of the data collection and drug safety signal mining agent for PMS according to an embodiment of the present invention. Figure 2 As shown, the intelligent agent includes:

[0181] The first unit is used to obtain historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping of drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by fusing drug molecular feature information and clinical medication information, wherein the multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes;

[0182] The second unit is used to use a heterogeneous graph neural network with a time series feature extraction function to perform representation learning on the multidimensional association network, construct training samples through a node type-aware sampling strategy, integrate drug action mechanism rules into the training process of the heterogeneous graph neural network, obtain a node vector representation of multidimensional information fusion, and construct a multi-task learning framework based on the node vector representation. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification;

[0183] The third unit is used to construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized drug adverse reaction data.

[0184] According to a third aspect of the embodiments of the present invention,

[0185] An electronic device is provided, comprising:

[0186] processor;

[0187] a memory for storing processor-executable instructions;

[0188] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0189] A fourth aspect of the embodiments of the present invention is:

[0190] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0191] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for data collection and drug safety signal mining for PMS, characterized in that: include: Acquire historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping on drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by integrating drug molecular feature information and clinical medication information, wherein the multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes; A heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, and training samples are constructed through a node type-aware sampling strategy. The drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network to obtain a node vector representation of multidimensional information fusion. Based on the node vector representation, a multi-task learning framework is constructed. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, and combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification; Construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized adverse drug reaction data.

2. The method according to claim 1, characterized in that The medical text pre-training model is used to standardize the mapping of drug names and adverse reaction descriptions in the original data to generate standardized adverse drug reaction data. Based on the standardized adverse drug reaction data, the drug molecular feature information and clinical medication information are integrated to construct a multi-layer drug knowledge graph, including: Inputting the original adverse drug reaction data into a medical text pre-training model, the medical text pre-training model performs word segmentation on the input text and extracts word vectors and term vectors, concatenating the word vectors with the term vectors to form a feature vector, inputting the feature vector into a bidirectional long short-term memory unit to extract context features, and performing sequence labeling through a conditional random field based on the context features to obtain initial drug entities and initial adverse reaction entities; Calculating the string similarity and semantic similarity between the initial drug entity and the standard medical term, determining the initial drug entity whose weighted combination value of the two similarities exceeds the mapping threshold as the standard drug entity, and mapping the initial adverse reaction entity to the standard adverse reaction entity to generate standardized drug adverse reaction data; The simplified linear input canonical representation of the drug molecule is converted into a molecular structure diagram with atom type and chemical bond type annotations. The molecular structure diagram is subjected to local feature extraction through multi-layer spatial convolution, and the extracted local features are aggregated through attention weights to obtain a drug molecule feature vector; Input clinical medication text into the bidirectional attention network, calculate the attention weight with standard medical terms, and obtain standardized clinical medication entities; The sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; the molecular similarity is calculated using drug molecular feature vectors to construct a molecular structure association network; and a clinical association network is constructed based on standardized clinical drug entities; Input the temporal association network, the molecular structure association network and the clinical association network into a hierarchical graph attention network, generate a node vector representation through a random walk algorithm, the node vector representation obtains multi-layer network features through a graph attention layer, and the multi-layer network features are input into a recurrent neural network to extract temporal features; The multi-layer network features and time series features are input into a heterogeneous graph neural network, and a multi-layer drug knowledge graph containing node vector representations is obtained through feature transformation training to form a multi-dimensional association network of drug information nodes and adverse reaction information nodes.

3. The method according to claim 2, characterized in that The sliding time window is used to detect the temporal co-occurrence relationship between standard drug entities and standard adverse reaction entities to construct a temporal association network; the molecular similarity is calculated using the drug molecular feature vector to construct a molecular structure association network; the clinical association network is constructed based on the standardized clinical drug entities, including: A sliding time window is used to detect the co-occurrence relationship between the standard drug entity and the standard adverse reaction entity in the time dimension to form a drug-adverse reaction entity pair, and a temporal co-occurrence metric value of the drug-adverse reaction entity pair is calculated for each time window. The temporal co-occurrence metric value is determined by the ratio of the number of co-occurrences of the drug-adverse reaction entity pair in the window to the total number of occurrences of each drug-adverse reaction entity pair; Performing a temporal significance test on the drug-adverse reaction entity pairs, generating a co-occurrence probability distribution by a random permutation method, establishing connections between the drug-adverse reaction entity pairs that pass the significance test, and using the temporal co-occurrence measurement value as a connection weight value to construct a temporal association network; Inputting the drug molecule feature vector into a similarity calculation model, the similarity calculation model performs attention weighted calculation on the chemical substructures in the drug molecule feature vector, outputs a molecular similarity matrix, inputs the molecular similarity matrix into a spectral clustering model for drug grouping, and connects drug molecules with similarities higher than a preset threshold according to the grouping results to construct a molecular structure association network; Forming a drug combination sequence according to the clinical drug entity, calculating the co-occurrence times of drugs in the drug combination sequence, obtaining the clinical co-occurrence support of the drug combination, establishing a Markov chain model for the drug combination sequence in time sequence, and calculating the transition probability between drugs in adjacent time sequence based on the Markov chain model; Constructing a bipartite graph of drugs and clinical indications, calculating the drug-clinical indication co-occurrence probability, calculating the mutual information ratio of the drug edge probability and the clinical indication edge probability according to the drug-clinical indication co-occurrence probability to obtain the drug-clinical indication association strength, and weightedly combining the clinical co-occurrence support, drug transfer probability and drug-clinical indication association strength as the connection weight value to construct a clinical association network; The connection weights in the temporal association network, the connection weights in the molecular structure association network, and the connection weights in the clinical association network are fused to construct a multi-layer drug association network.

4. The method according to claim 1, characterized in that: A heterogeneous graph neural network with a time series feature extraction function is used to perform representation learning on the multidimensional association network, a node type-aware sampling strategy is used to construct training samples, and the drug action mechanism rules are integrated into the training process of the heterogeneous graph neural network to obtain a node vector representation of multidimensional information fusion, including: Receiving a multidimensional association network and a drug action mechanism rule, performing a linear transformation on a node feature vector in the multidimensional association network, and generating a query vector, a key vector, and a value vector; An attention score matrix is ​​obtained by performing a dot product operation on the query vector and the key vector, the attention score matrix is ​​normalized to obtain a temporal attention weight, a matrix multiplication operation is performed on the temporal attention weight and the value vector to obtain a temporal context vector, and the temporal context vector is concatenated with the node feature vector and then the temporal enhancement feature is obtained through a feedforward neural network; The time series enhancement feature is weightedly summed with the edge weight in the multidimensional association network to obtain the edge weight of the fused time series feature, the degree centrality of each node is calculated as the node importance score according to the edge weight of the fused time series feature, the nodes are divided into different categories according to the node type, and probability sampling is performed in each category according to the node importance score to obtain a training node set; Extracting drug nodes associated with the mechanism of action from the drug action mechanism rules as drug node pairs, calculating the cosine similarity of the drug node pairs in the training node set, and taking the difference between the similarity and a preset node similarity threshold as the rule loss; Input the training node set into the heterogeneous graph neural network for node classification, use cross entropy to calculate the classification loss, and use the weighted sum of the classification loss and the rule loss as the total loss function; Based on the total loss function, the parameters of the heterogeneous graph neural network are gradient updated, and the iteration is repeated until the preset number of iterations is met to obtain the final node vector representation of multi-dimensional information fusion.

5. The method according to claim 1, characterized in that Based on the node vector representation, a multi-task learning framework is constructed. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification, including: Obtaining and concatenating the drug node vector and the adverse reaction node vector in the node vector representation, extracting features through a multi-layer attention-enhanced feedforward neural network with residual connections to generate a drug-adverse reaction shared feature vector; Constructing a time series feature sequence matrix, wherein the rows of the time series feature sequence matrix represent different time windows, and the columns represent feature sets corresponding to the drug node vector and the adverse reaction node vector in each time window, wherein the feature set includes statistical association features calculated based on an adaptive time window, drug molecular structure features, and drug target features; Inputting the temporal feature sequence matrix into a bidirectional long short-term memory network provided with a gating mechanism, wherein the gating mechanism adaptively adjusts the feature weight according to the importance of the time window to generate a temporal context vector; A multi-head attention mechanism is used to fuse the drug-adverse reaction shared feature vector with the temporal context vector to obtain a drug-adverse reaction multi-task feature vector, and the drug-adverse reaction multi-task feature vector is input into an association prediction network and a risk prediction network respectively to generate an initial drug-adverse reaction association probability and an initial risk level probability; Calculate the binary cross entropy loss based on the initial drug-adverse reaction association probability, calculate the multi-category cross entropy loss based on the initial risk level probability, construct a multi-task loss function in combination with the time series smoothing constraint, update the model parameters by iteratively optimizing the multi-task loss function, and obtain the optimized drug-adverse reaction association probability and risk level probability; The kernel density estimation method is used to calculate the probability distribution of the current time window based on the optimized drug-adverse reaction association probability, determine the dynamic association threshold, and screen candidate drug safety signals from the optimized drug-adverse reaction association probability and risk level probability according to the dynamic association threshold and the preset risk threshold.

6. The method according to claim 1, characterized in that Constructing a hierarchical evaluation system including a signal strength dimension and a clinical relevance dimension, verifying the candidate drug safety signal based on the hierarchical evaluation system, confirming the reliability of the candidate drug safety signal in combination with clinical data, outputting a target drug safety signal, mapping the target drug safety signal to the hierarchical evaluation system, determining an early warning level based on a mapping position, setting a differentiated monitoring period and a trigger threshold according to the early warning level, and inputting the node vector representation into an early warning analysis model, and performing real-time risk early warning on the standardized adverse drug reaction data, including: Construct a hierarchical evaluation system of signal strength dimension and clinical relevance dimension. The signal strength dimension is based on the strength of drug-adverse reaction causal relationship, the number of adverse reaction reports, and the time trend of adverse reactions to construct a weighted scoring model to obtain a signal strength score. The clinical relevance dimension is based on the severity of adverse reactions, population sensitivity, and clinical interventionability to construct a hierarchical analysis network to obtain a clinical relevance score. Inputting the candidate drug safety signal, signal strength score, and clinical relevance score into the screening model to obtain a first screening result, inputting the first screening result into a cohort analysis unit, a literature analysis unit, and a social media analysis unit, respectively, and combining the verification results of each unit using an integrated learning model to obtain a multi-source verification result; Inputting the drug molecular characteristics, metabolic information, action network information and the multi-source verification results corresponding to the candidate drug safety signal into a graph neural network model, and determining the target drug safety signal based on the output results of the graph neural network model; Inputting the target drug safety signal, the signal strength score, and the clinical relevance score into a pre-trained deep reinforcement learning model to obtain a warning level; A monitoring strategy is constructed according to the warning level, and the monitoring strategy and the time series characteristics of the target drug safety signal are input into a hybrid neural network model. The time series characteristics include the number characteristics of adverse reaction reports, the cumulative proportion characteristics and the geographical distribution characteristics. Based on the output results of the hybrid neural network model, risk warnings are issued for adverse drug reaction data.

7. The method according to claim 6, characterized in that The target drug safety signal, the signal strength score and the clinical relevance score are input into a pre-trained deep reinforcement learning model to obtain the warning levels including: Collect characteristic information of the target drug safety signal, the characteristic information including signal structured representation, signal intensity temporal variation characteristics, clinical relevance score characteristics and historical warning records, construct the characteristic information into a state space and map it into a state vector; Constructing an action space for warning level adjustment, the action space including a warning level raising action, a warning level lowering action and a warning level maintaining action, mapping the action space into an action vector, and the action vector and the state vector form a state-action pair; Constructing a reward function for early warning evaluation, calculating a weighted reward value based on an early warning accuracy index, an early warning timeliness index, and a supervision resource consumption index, and forming a training sample with the state-action pair and the weighted reward value; Constructing a dual learning network structure including a current network and a target network, storing the training samples in an experience replay pool, sampling from the experience replay pool to obtain a state transition sequence, inputting the state transition sequence into the dual learning network structure, and using a temporal difference algorithm to alternately optimize and update the parameters of the current network and the target network; A decreasing greedy strategy is used for action selection. The corresponding action is selected based on the current state and environmental feedback is obtained. The environmental feedback is input into the dual learning network structure to obtain the optimal early warning decision strategy. The real-time feature vector of the target drug safety signal is input into the optimal early warning decision strategy, and the corresponding early warning level is output.

8. A data collection and drug safety signal mining agent for PMS, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical monitoring data in the PMS system as raw data, use a medical text pre-training model to perform standardized mapping of drug names and adverse reaction descriptions in the raw data, generate standardized adverse drug reaction data, and build a multi-layer drug knowledge graph based on the standardized adverse drug reaction data by fusing drug molecular feature information and clinical medication information, wherein the multi-layer drug knowledge graph forms a multi-dimensional association network including drug information nodes and adverse reaction information nodes; The second unit is used to use a heterogeneous graph neural network with a time series feature extraction function to perform representation learning on the multidimensional association network, construct training samples through a node type-aware sampling strategy, integrate drug action mechanism rules into the training process of the heterogeneous graph neural network, obtain a node vector representation of multidimensional information fusion, and construct a multi-task learning framework based on the node vector representation. The multi-task learning framework simultaneously completes the drug-adverse reaction association probability prediction task and the risk level prediction task, combines the time series analysis method to detect the dynamic change characteristics of the association probability and the risk level, and generates a candidate drug safety signal with a risk level identification; The third unit is used to construct a hierarchical evaluation system including signal strength dimension and clinical relevance dimension, verify the candidate drug safety signal based on the hierarchical evaluation system, confirm the reliability of the candidate drug safety signal in combination with clinical data, output the target drug safety signal, map the target drug safety signal to the hierarchical evaluation system, determine the warning level based on the mapping position, set differentiated monitoring periods and trigger thresholds according to the warning level, and input the node vector representation into the warning analysis model to perform real-time risk warning on the standardized drug adverse reaction data.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images