Drug safety multi-center joint evaluation method and system based on graph neural network and federated learning

Abstract

The invention relates to the technical field of drug safety evaluation, in particular to a multi-center combined evaluation method and system for drug safety based on a graph neural network and federated learning. Known and unknown drug interaction is systematically predicted based on a drug multi-relation knowledge graph and a graph neural network, a key path of DDI is identified through a graph attention mechanism, a molecular mechanism of the interaction is revealed, and natural language interpretation is generated. And meanwhile, through privacy protection and multi-center cooperation, a federal learning architecture is utilized to break data islands and improve the external effectiveness of an evaluation conclusion on the premise of protecting patient privacy and meeting data compliance requirements. The heterogeneity of data of different mechanisms is effectively evaluated through distribution deviation detection, deviation caused by blind extrapolation is avoided, a real-world evidence methodology report and a data traceability auditing clue are automatically generated, the requirement of a supervision mechanism for real-world evidence quality is met, medicine supervision decision is supported, and medicine research, development and review are accelerated.

Description

Technical Field

[0001] This invention relates to the field of drug safety evaluation technology, and in particular to a multi-center joint evaluation method and system for drug safety based on graph neural networks and federated learning. Background Technology

[0002] Post-marketing safety monitoring of drugs is crucial for ensuring public medication safety and supporting drug re-registration and indication expansion. Drug-drug interactions (DDIs) are an important aspect of drug safety evaluation; multiple drug combinations may lead to enhanced or diminished efficacy or the development of new adverse reactions. With an aging population and rising prevalence of chronic diseases, the simultaneous use of multiple medications by patients is becoming increasingly common, highlighting the growing importance of DDI monitoring.

[0003] Traditional disease-drug interaction (DDI) monitoring relies primarily on spontaneous reporting systems (SRS) and literature reviews, but these methods suffer from limitations such as incomplete reporting, high underreporting rates, and difficulty in establishing causal relationships. Clinical trials typically exclude patients using multiple drugs, making it difficult to comprehensively assess the risk of DDIs in the real world. Real-world data (RWD), derived from electronic medical records and health insurance databases, covers a large number of patients using multiple drugs, providing a new data source for DDI monitoring. However, existing methods still have shortcomings in systematically identifying DDIs, predicting unknown interactions, and explaining interaction mechanisms.

[0004] In recent years, knowledge graphs and graph neural networks (GNNs) have shown great potential in drug discovery and safety evaluation. Drug knowledge graphs integrate heterogeneous data from multiple sources, including drug-target, drug-disease, and drug-side effects, providing rich background knowledge for drug discovery and safety assessment (DDI) prediction. Graph neural networks, capable of effective representation learning and relationship prediction on graph-structured data, have been successfully applied to DDI prediction tasks.

[0005] However, data from a single institution often has limited sample size and insufficient representativeness of the patient population, making it difficult to support large-scale DDI monitoring. Multi-center collaboration across institutions can expand sample size and improve the external validity of evaluation conclusions, but it faces challenges related to data privacy protection and data silos. Patient data from various medical institutions involves sensitive privacy information and is subject to strict legal and regulatory constraints, making direct sharing of raw data impossible.

[0006] Federated learning, an emerging distributed machine learning technology, enables collaborative modeling among multiple parties without sharing the original data. Each participant trains its model on local data, only uploading updates to the model parameters. The global model is then aggregated through a central server, thus achieving collaborative learning while protecting data privacy.

[0007] Furthermore, drug regulatory agencies impose stringent quality requirements on the acceptance of Real-World Evidence (RWE). Regulatory bodies such as the FDA and NMPA have issued guidelines for RWE application, clarifying quality assessment dimensions such as data relevance, data reliability, study design rationality, and appropriateness of statistical analysis. Existing DDI monitoring systems often lack standardized methodological documentation and data traceability, making it difficult to meet regulatory compliance requirements.

[0008] To address the above issues, there is an urgent need in this field for an innovative drug safety evaluation method that can: (1) perform systematic DDI prediction based on drug knowledge graphs and graph neural networks to identify known and unknown drug interactions; (2) provide interpretable interaction mechanisms to enhance clinical credibility; (3) achieve cross-institutional privacy-preserving multi-center collaboration through federated learning; and (4) generate real-world evidence reports that meet regulatory requirements to support drug registration and post-marketing evaluation. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning, comprising the following steps: S1. Construct a multi-relationship knowledge graph for drugs based on a multi-source public database of drugs; S2. Based on the drug multi-relationship knowledge graph, construct a graph neural network model and perform representation learning to obtain the embedding vector of each node; S3. Based on the embedding vectors of each node, predict the drug-drug interactions to obtain the probability of drug-drug interactions and the types of drug-drug interactions; S4. Calculate the safety score of the candidate drug regimen based on the corresponding weights of the drug interaction types and the probability of drug interaction occurrence; S5. Based on the safety score, the contribution of each edge of the graph neural network to the drug interaction prediction results is calculated using the graph attention algorithm, and an interpretable description of the drug interaction mechanism is generated by combining it with the natural language description algorithm. S6. Train the graph neural network model in multiple medical centers based on local real-world data using steps S1-S5, calculate the gradient of the model parameters for each medical center, encrypt and upload the gradient of the model parameters for each medical center using a federated learning framework, obtain the global graph neural network model parameters for multiple centers, and update the graph neural network model for each medical center until the model converges. S7. Based on the parameters of the multi-center global model, calculate the distribution shift difference value of the explainable drug interaction mechanism description of each medical center through the covariate distribution test algorithm. When the distribution shift difference value exceeds the preset threshold, it indicates that there is uncertainty in the explainable drug interaction mechanism description of the current medical center for the current patient, generates a drug safety evaluation report and issues an early warning.

[0010] Based on the above methods, the present invention also provides a multi-center joint evaluation system for drug safety based on graph neural networks and federated learning. The system is implemented based on any one of the above-mentioned multi-center joint evaluation methods for drug safety based on graph neural networks and federated learning, including a knowledge graph construction module for constructing a drug multi-relationship knowledge graph based on a multi-source public database of drugs. The graph neural network construction module is used to construct a graph neural network model and perform representation learning based on the drug multi-relationship knowledge graph to obtain the embedding vector of each node; The interaction prediction module is used to predict drug interactions based on the embedding vectors of each node, and to obtain the probability of drug interaction and the type of drug interaction. The safety assessment module calculates the safety score of the candidate drug regimen based on the corresponding weights of the types of drug interactions and the probability of drug interactions occurring. An interpretable mechanism generation module is used to calculate the contribution of each edge of the graph neural network to the drug interaction prediction results based on the safety score using a graph attention algorithm, and to generate an interpretable description of the drug interaction mechanism by combining it with a natural language description algorithm. The federated learning coordination module is used to train the graph neural network model in multiple medical centers based on local real-world data using steps S1-S5, calculate the gradient of the model parameters for each medical center, encrypt and upload the gradient of the model parameters for each medical center through the federated learning framework, obtain the global graph neural network model parameters for multiple centers, and update the graph neural network model for each medical center until the model converges. The distribution shift detection and early warning module is used to calculate the distribution shift difference value of the explainable drug interaction mechanism description of each medical center based on the multi-center global model parameters and the covariate distribution test algorithm. When the distribution shift difference value exceeds the preset threshold, it indicates that there is uncertainty in the explainable drug interaction mechanism description of the current medical center for the current patient, generates a drug safety evaluation report and issues an early warning.

[0011] The beneficial effects of this invention are that, by using a drug multi-relationship knowledge graph and graph neural network, it can systematically predict known and unknown drug interactions, identify potential safety risks, and compensate for the shortcomings of traditional methods in assessing drug interactions (DDIs). Furthermore, it identifies critical paths of DDIs through graph attention mechanisms, reveals the molecular mechanisms of interactions, generates natural language explanations, and enhances clinicians' understanding and trust in the prediction results. On the other hand, in addressing the data silo effect, it achieves joint modeling of cross-institutional data through privacy protection and multi-center collaboration using a federated learning architecture. While protecting patient privacy and meeting data compliance requirements, it breaks down data silos, expands the sample size, and improves the external validity of evaluation conclusions. Distribution offset detection effectively assesses the heterogeneity of data from different institutions, indicating the transportability risk of evaluation conclusions, avoiding bias caused by blind extrapolation, and improving the scientific rigor of evaluation conclusions. The automatically generated real-world evidence methodology report and data source audit clues meet the requirements of regulatory agencies such as the FDA and NMPA for the quality of real-world evidence, support drug regulatory decisions, and accelerate drug development and review processes. On the other hand, the drug knowledge graph integrates multi-source heterogeneous data, covering tens of thousands of drugs, thousands of targets and diseases, providing rich background knowledge for DDI prediction and improving prediction accuracy. Furthermore, it adopts a modular system architecture, facilitating the expansion to new data sources and new model algorithms. The graph neural network model has strong generalization ability and can predict new DDIs not present in the training data. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning, according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the structure of a multi-center joint evaluation system for drug safety based on graph neural networks and federated learning, according to an embodiment of the present invention.

[0014] Figure 3 This is a schematic diagram of the terminal device structure of a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning, according to an embodiment of the present invention.

[0015] Figure 4This is a schematic diagram of a computer-readable storage medium structure for a multi-center joint evaluation method of drug safety based on graph neural networks and federated learning, according to an embodiment of the present invention.

[0016] In the diagram, 200 is the terminal device, 210 is the memory, 211 is the RAM, 212 is the cache memory, 213 is the ROM, 214 is the program / utility, 215 is the program module, 220 is the processor, 230 is the bus, 240 is the external device, 250 is the I / O interface, 260 is the network adapter, and 300 is the program product. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0018] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0019] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0020] Example 1: like Figure 1 As shown, Embodiment 1 of the present invention provides a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning, including the following steps: S1. Construct a multi-relationship knowledge graph for drugs based on a multi-source public database of drugs; S2. Based on the drug multi-relationship knowledge graph, construct a graph neural network model and perform representation learning to obtain the embedding vector of each node; S3. Based on the embedding vectors of each node, predict the drug-drug interactions to obtain the probability of drug-drug interactions and the types of drug-drug interactions; S4. Calculate the safety score of the candidate drug regimen based on the corresponding weights of the drug interaction types and the probability of drug interaction occurrence; S5. Based on the safety score, the contribution of each edge of the graph neural network to the drug interaction prediction results is calculated using the graph attention algorithm, and an interpretable description of the drug interaction mechanism is generated by combining it with the natural language description algorithm. S6. Train the graph neural network model in multiple medical centers based on local real-world data using steps S1-S5, calculate the gradient of the model parameters for each medical center, encrypt and upload the gradient of the model parameters for each medical center using a federated learning framework, obtain the global graph neural network model parameters for multiple centers, and update the graph neural network model for each medical center until the model converges. S7. Based on the parameters of the multi-center global model, calculate the distribution shift difference value of the explainable drug interaction mechanism description of each medical center through the covariate distribution test algorithm. When the distribution shift difference value exceeds the preset threshold, it indicates that there is uncertainty in the explainable drug interaction mechanism description of the current medical center for the current patient, generates a drug safety evaluation report and issues an early warning.

[0021] Furthermore, step S1 includes the following sub-steps: S101. Extract drug relationship data from a multi-source public drug database. The drug relationship data includes drug target relationship data, basic drug information data, drug side effect relationship data, drug pathway relationship data, drug-drug interaction relationship data, and protein-protein relationship data. S102. After standardizing the extracted drug relationship data, the drug entities are mapped to a unified drug coding system, and entity disambiguation and fusion are performed on different representations of the same drug entity. S103. Based on the drug coding system after entity disambiguation and fusion, a heterogeneous graph representation is constructed to obtain a drug multi-relationship knowledge graph, represented as follows: ; Where V represents the set of nodes, E represents the set of edges, and R represents the set of relation types.

[0022] Specifically, the implementation flow of each sub-step in the above embodiments is as follows: First, drug relationship data is extracted from multiple publicly available drug databases, including DrugBank (drug-target relationship data, basic drug information data), SIDER (drug-side effect relationship data), OMIM (disease-gene relationship data), STRING (protein-protein interaction data), KEGG (drug-metabolic pathway relationship data), and TWOSIDES (drug-drug interaction data). The extracted relationship types include, but are not limited to: drug-target binding, drug-side effect association, drug-drug interaction, drug-disease treatment, protein-protein interaction, gene-disease association, and drug-metabolic enzyme inhibition / induction.

[0023] Next, the extracted relational data is standardized: drug entities are mapped to a unified drug coding system, such as DrugBank ID or ATC coding; protein entities are mapped to UniProtID; disease entities are mapped to ICD-10 coding or OMIM coding; and side effect entities are mapped to MedDRA terms. For different representations of the same entity (such as generic drug name, brand name, and chemical name), entity disambiguation and fusion are performed to ensure the consistency of the knowledge graph.

[0024] Then, a heterogeneous graph representation is constructed. The knowledge graph is represented as G=(V,E,R), where V is the set of nodes, including drug nodes, protein nodes, disease nodes, and side effect nodes; E is the set of edges, representing the relationships between entities; and R is the set of relationship types. Different types of nodes have different feature dimensions. For example, drug nodes may contain features such as molecular fingerprints and physicochemical properties, while protein nodes may contain features such as structural domain information.

[0025] Finally, quality control is performed on the knowledge graph. This includes removing redundant relationships (such as duplicate records of the same pair of entities); correcting erroneous mappings (such as incorrect drug-target associations); supplementing missing key relationships (such as mining DDIs not yet included in the database from the latest literature); and using methods such as expert review or literature verification to improve the accuracy and completeness of the knowledge graph.

[0026] Furthermore, step S2 includes the following sub-steps: S201: Based on the drug multi-relationship knowledge graph, embedding learning model of the drug multi-relationship knowledge graph is performed through a relation graph convolutional network; S202: Based on the message passing mechanism, the neighbor information of nodes in the relation graph convolutional network is distinguished and aggregated according to the drug relationship type, and the node embedding is updated. The node embedding update formula is expressed as: ; in, Represents a node In the Layer embedding vectors, Represents a node In the Layer embedding vectors, Indicates the type of drug relationship. Represents the set of drug relationship types. Indicates a relationship through medication With nodes The set of connected neighbor nodes, and Let represent the learnable weight matrices, Indicates the activation function; S203: Repeat step S202. After multiple layers of message passing, the embedding vector of each node is fused with the structural and feature information in the multi-hop neighborhood to obtain the embedding vector of each node.

[0027] Specifically, the implementation flow of each sub-step in the above embodiments is as follows: First, based on the drug multi-relational knowledge graph, neighbor information of nodes is aggregated through a message passing mechanism. A Relational Graph Convolutional Network (R-GCN) or Graph Attention Network (GAT) architecture is adopted to support modeling of multiple relation types in heterogeneous graphs. Specifically, in the method described in this embodiment, a graph neural network is used to learn the representation of the drug knowledge graph, employing a Relational Graph Convolutional Network (R-GCN) architecture. R-GCN aggregates neighbor information of nodes through a message passing mechanism, learning the low-dimensional embedding representation of nodes. After multiple layers of message passing, the embedding vector of each node fuses the structural and feature information within its multi-hop neighborhood, thereby capturing the complex association patterns between drugs, targets, diseases, and side effects. The node embedding update formula is: ; in, Represents a node In the Layer embedding vectors, Represents a node In the Layer embedding vectors, Indicates the type of drug relationship. Represents the set of drug relationship types. Indicates a relationship through medication With nodes The set of connected neighbor nodes, and Let represent the learnable weight matrices, This represents the activation function; then, after multiple layers of message passing, the embedding vector of each node incorporates the structural and feature information from its multi-hop neighborhood.

[0028] Furthermore, step S3, predicting drug-drug interactions, includes using a bilinear decoder to perform interactive calculations on the node embedding vectors of any two drugs i and j in the candidate dosing regimens, predicting the probability of drug-drug interactions. The interactive calculation formula is expressed as follows: ; in, Indicates the probability of an interaction occurring. and These represent drug nodes, Indicates the type of interaction relationship. and Let represent the node embedding vectors, and T represent the matrix transpose. The weight matrix representing the specific interaction relationship is normalized to the probability of the interaction occurring using the sigmoid function, as shown in the formula: ; ; Here, DDI stands for Drug-Drug Interaction. This indicates the type of interaction between drug i and drug j. The probability value of interaction occurring. This represents the sigmoid function. Indicates that the drugs are compatible. The total number of drugs is represented by the interaction prediction of all drug pairs in the candidate drug regimen, which yields the probability of drug i and drug j occurring in each drug-drug interaction type.

[0029] Specifically, in step S3, drug-drug interactions are predicted based on node embedding vectors. For any two drugs, a bilinear decoder is used to calculate the interaction score, and the probability of interaction occurrence is obtained by normalization using the sigmoid function. This method can not only predict known drug interactions (DDIs) but also discover new DDIs that have not appeared in the training data, demonstrating strong generalization ability.

[0030] Furthermore, step S4, which calculates the safety score of the candidate drug regimen, includes assigning a severity weight to each interaction type. The safety score calculation formula is expressed as: ; in, Indicates the safety score. This represents a set of drug-drug relationship types, where DDI represents drug-drug interactions. This indicates the type of interaction between drug i and drug j. The probability value of interaction occurring. The severity weight for interaction type r is determined based on a pre-defined clinical severity grading.

[0031] Specifically, different types of drug interactions (DDIs) have different clinical severity levels, such as serious interactions (potentially life-threatening), moderate interactions (requiring monitoring or dose adjustment), and mild interactions (usually requiring no treatment). Based on the clinical severity grading, a weighting of 1.0 can be assigned to serious interactions, 0.5 to moderate interactions, and 0.2 to mild interactions; a negative sign indicates that more interactions result in a lower safety score. This score reflects the overall DDI risk burden of the candidate drug regimen.

[0032] Furthermore, step S5 includes the following sub-steps: S501: Based on the safety score, the attention weights of each edge of the graph neural network are calculated through the graph attention mechanism to obtain the maximum attention weight, and the corresponding edge is identified as the propagation path that contributes the most between drugs. S502: Based on the propagation path obtained in step S501, extract the intermediate nodes and drug relationship types on the path, and generate an interpretable description of the drug interaction mechanism based on the propagation path information using a natural language description algorithm.

[0033] Specifically, the implementation process of each sub-step in the above embodiments is as follows: First, the attention weight of each edge is calculated using the graph attention mechanism to identify the propagation path that contributes the most between drugs; then, the intermediate nodes and relationships on the path are extracted, such as: drug A → target protein X → protein interaction → target protein Y → drug B; then, a natural language description is generated based on the path information to explain the molecular mechanism of the interaction; finally, the contribution of each path to the interaction prediction is quantified to provide interpretability evidence.

[0034] Specifically, in the method described in this embodiment of the invention, step S5 identifies key pathways of drug-drug interactions. Using graph attention mechanisms, the contribution of each edge in the drug knowledge graph to the prediction of drug interaction differences (DDIs) can be calculated. By extracting the path with the highest contribution, the molecular mechanism of DDIs can be revealed. For example, two drugs may interact through mechanisms such as common targets, metabolic enzyme competition, and protein interactions. Based on the path information, the system automatically generates natural language descriptions to provide clinicians with understandable explanations.

[0035] Furthermore, step S6 includes the following sub-steps: S601: Deploy graph neural network models in local training environments of multiple medical centers, and repeat steps S1-S5 based on local real-world data to train the models and obtain gradient updates of graph neural network model parameters for each medical center. S602: Based on a pre-set secure multi-party computation protocol, each medical center encrypts the gradient updates of the graph neural network model parameters and transmits them to the central aggregation server. S603: Based on the gradient update of the encrypted graph neural network model parameters uploaded by each medical center, the parameters of the global graph neural network model of the multi-center are generated by federated average aggregation through the central aggregation server. S604: Update the local model based on the global model parameters downloaded from each medical center; S605: Repeat steps S602 to S604 until the model converges.

[0036] Specifically, the implementation process of each sub-step in the above embodiment is as follows: First, deploy the graph neural network model in the local environment of each participating institution; then, each institution trains the model based on local real-world data and calculates the gradient update of the local model parameters; next, through a secure multi-party computation protocol, each institution uploads the encrypted model parameter gradient to the central aggregation server; then, the central aggregation server performs federated averaging aggregation on the encrypted gradient to generate global model parameters; finally, each institution downloads the global model parameters to update its local model until the model converges.

[0037] Specifically, in the method described in this embodiment of the invention, step S6 employs a federated learning approach for privacy-preserving multi-center collaboration. Each participating institution deploys its model in its local environment, trains it on local data, and calculates gradient updates for the model parameters. Through a secure multi-party computation protocol (such as homomorphic encryption or secret sharing), each institution uploads its encrypted gradients to a central aggregation server. The central server calculates the aggregated gradients and updates the global model parameters without decrypting the gradients of individual institutions. Each institution downloads the global model parameters and updates its local model. This process is repeated until the model converges. This method ensures that the original patient data does not leave the medical institution, meeting data privacy protection and compliance requirements.

[0038] Specifically, based on a pre-set secure multi-party computation protocol, each medical center encrypts the gradient updates of the graph neural network model parameters and transmits them to the central aggregation server. This includes using secure aggregation protocols, such as homomorphic encryption or secret sharing, to protect the privacy of gradient information. Specifically, each institution manages its local gradients... Encryption is performed to obtain the encrypted local gradients for each medical institution. ,in Indicates the medical institution number. This indicates an encryption operation, which is then uploaded to a central server. The central server can only compute the ciphertext of the aggregated gradients without decrypting the gradients of individual institutions. The central aggregation server performs federated averaging on the encrypted gradients to generate global model parameters. Therefore, the federated averaging algorithm calculates a weighted average of the gradients from each institution, with each institution averaging its local gradient. Encryption is performed to obtain ciphertext. The ciphertext formula for calculating the aggregate gradient in the ciphertext domain by the central server is expressed as: ; in, Indicates the medical institution number. Indicates medical institutions The number of samples, This represents the total number of samples. Indicates medical institutions The sample size proportion is used as a weighting coefficient when applying the federal average to avoid abnormal global model parameters caused by an excessively small sample size in medical institutions (k). This indicates that the central server decrypts the global gradient and updates the global model parameters.

[0039] Furthermore, step S7 includes the following sub-steps: S701: Based on the parameters of the multicenter global model, the local patient population characteristics of each medical center are quantitatively analyzed through the covariate distribution test algorithm to obtain the distribution differences of patient population characteristics among different medical centers. S702: Based on the differences in the distribution of patient population characteristics among different medical centers, compare the model performance indicators of local real-world data of each medical center to obtain the distribution offset difference value of the interpretable drug interaction mechanism explanation of each medical center. S703: If the distribution offset difference value exceeds the preset threshold, it indicates that there is uncertainty regarding the explainable drug interaction mechanism of the current medical center for the current patient, and issues an early warning.

[0040] Furthermore, the drug safety evaluation report in step S7 includes a description of the data source, bias control methods, missing data processing strategies and cohort construction process, and data traceability audit clues; the description of the data source includes the medical center number, data time range, and patient population characteristics; the bias control methods include selective bias control methods and confounding bias control methods; and the data traceability audit clues include a complete processing flow record from step S1 to step S7.

[0041] Specifically, due to differences in patient population characteristics, disease spectrum, and treatment guidelines among different medical institutions, data distribution may be inconsistent. A covariate distribution test algorithm is used to quantify the degree of heterogeneity among institutions. If the heterogeneity is large, the generalization performance of the global model may be poor in some institutions, and the applicability of the evaluation conclusions to patients in those institutions may be uncertain. The system generates a transportability risk warning to remind clinicians to interpret the evaluation results with caution.

[0042] Specifically, the implementation process of each sub-step in the above embodiments is as follows: First, covariate distribution testing methods, such as the Kolmogorov-Smirnov test or the Maximum Mean Discrepancy test, are used to quantify the distribution differences of patient population characteristics among different institutions; then, the model performance indicators of the local validation sets of each institution are compared to evaluate the generalization ability of the model in different institutions; if the distribution deviation exceeds a preset threshold, a regional extrapolation warning is generated, indicating that there is uncertainty in the applicability of the evaluation conclusion to patients in a specific institution. Finally, the data source description is recorded, including participating institutions, data time range, patient population characteristics, inclusion and exclusion criteria, bias control methods, including strategies for handling selection bias and confounding bias, missing data processing strategies and cohort construction process are recorded, data traceability audit clues are established, the complete processing flow from raw data to evaluation conclusions is recorded, the methodology report fields are mapped to the quality dimension requirements of drug regulatory agencies for real-world data and real-world evidence, and standardized evidence report documents that meet the regulatory requirements of FDA / NMPA and other regulatory bodies are automatically generated.

[0043] Specifically, the standardized evidence report document includes sections on data source descriptions, study design, bias control methods, statistical analysis methods, and results presentation, meeting the quality requirements for RWE (real-world studies) in drug regulatory standards. A data traceability audit trail is established, recording the complete processing flow from raw data to evaluation conclusions to ensure the repeatability and auditability of the evaluation process.

[0044] Example 2

[0045] like Figure 2 As shown in Example 1, Example 2 of this invention proposes a multi-center joint evaluation system for drug safety based on graph neural networks and federated learning, which is implemented based on a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning.

[0046] Specifically, the system includes a knowledge graph construction module, which is used to construct a drug multi-relationship knowledge graph based on a multi-source public drug database; The graph neural network construction module is used to construct a graph neural network model and perform representation learning based on the drug multi-relationship knowledge graph to obtain the embedding vector of each node; The interaction prediction module is used to predict drug interactions based on the embedding vectors of each node, and to obtain the probability of drug interaction and the type of drug interaction. The safety assessment module calculates the safety score of the candidate drug regimen based on the corresponding weights of the types of drug interactions and the probability of drug interactions occurring. An interpretable mechanism generation module is used to calculate the contribution of each edge of the graph neural network to the drug interaction prediction results based on the safety score using a graph attention algorithm, and to generate an interpretable description of the drug interaction mechanism by combining it with a natural language description algorithm. The federated learning coordination module is used to train the graph neural network model in multiple medical centers based on local real-world data using steps S1-S5, calculate the gradient of the model parameters for each medical center, encrypt and upload the gradient of the model parameters for each medical center through the federated learning framework, obtain the global graph neural network model parameters for multiple centers, and update the graph neural network model for each medical center until the model converges. The distribution shift detection and early warning module is used to calculate the distribution shift difference value of the explainable drug interaction mechanism description of each medical center based on the multi-center global model parameters and the covariate distribution test algorithm. When the distribution shift difference value exceeds the preset threshold, it indicates that there is uncertainty in the explainable drug interaction mechanism description of the current medical center for the current patient, generates a drug safety evaluation report and issues an early warning.

[0047] Furthermore, the digital twin model construction module also includes: a multimodal preprocessing unit for feature extraction and encoding of structured clinical indicator data, medical order data, imaging test summary data, and free text data; a temporal modeling unit for constructing a temporal state evolution model based on the Transformer architecture; a generative decoding unit for generating patient state evolution trajectories under different medication regimens; and a counterfactual reasoning unit for calculating efficacy indicator predictions and adverse event risk distributions.

[0048] Furthermore, the causal assessment module also includes: a data playback unit for extracting medication trajectory data from a historical real-world database; a propensity score estimation unit for estimating the probability of choosing a specific medication action under different states; a value function estimation unit for predicting the expected long-term return of state-action pairs; a dual robust estimation unit for combining propensity scores and value function estimation to perform off-policy assessment; and a confidence interval calculation unit for calculating the lower confidence bound of the policy value and the upper bound of the policy risk.

[0049] Specifically, the workflow of this system is as follows: First, the knowledge graph construction module acquires information on drugs, targets, diseases, side effects, and their associations from a multi-source public drug database. It integrates, cleans, and structures the multi-source data to construct a multi-relationship knowledge graph containing various entities and relationship types. Then, the graph neural network construction module builds a graph neural network model based on the multi-relationship knowledge graph. It learns the representation of nodes in the knowledge graph through node feature encoding and a multi-layer message passing mechanism to obtain the embedding vectors of each node. Next, the interaction prediction module uses the node embedding vectors to perform interaction calculations on any two drugs in the candidate drug regimens, predicting the probability and type of interaction between the drugs using a decoder. Subsequently, the safety assessment module calculates the comprehensive safety score of the candidate drug regimens based on the weights corresponding to the interaction types and the probability of interaction. Finally, the interpretable mechanism generation module calculates the contribution of each edge in the graph neural network to the drug interaction prediction results based on the safety score using a graph attention algorithm. The system first generates an interpretable drug interaction mechanism description using natural language processing algorithms. Then, the federated learning coordination module executes training processes for knowledge graph construction, graph neural network representation learning, interaction prediction, safety assessment, and interpretable mechanism generation in local real-world data environments across multiple medical centers. It calculates the gradients of local model parameters and, through the federated learning framework, encrypts, uploads, and aggregates these gradients to generate multi-center global graph neural network model parameters. These parameters are then synchronized to the local models in each medical center for iterative updates until convergence. Finally, the distribution offset detection and early warning module calculates the distribution offset difference value of the interpretable drug interaction mechanism descriptions generated by each medical center based on the final multi-center global model parameters using a covariate distribution test algorithm. When the difference value exceeds a preset threshold, it indicates uncertainty regarding the interpretable drug interaction mechanism description for the target patient in the current medical center and outputs a drug safety evaluation report with an early warning.

[0050] Example 3

[0051] Based on Example 1, there is a multi-center joint evaluation application scenario for drug safety. This scenario adopts the multi-center joint evaluation and system for drug safety based on graph neural networks and federated learning as described in the above examples. This example demonstrates a joint evaluation across 5 hospitals of the efficacy and cardiovascular safety of second-line hypoglycemic drugs (SGLT-2 inhibitors vs. DPP-4 inhibitors) for patients with type 2 diabetes.

[0052] First, the federated learning architecture was deployed: five hospitals participated in the joint evaluation, with each hospital deploying a local client and running the graph neural network DDI evaluation model on its local server. A central coordination server was deployed on the private cloud of a medical research institution, responsible for coordinating the federated learning process.

[0053] Local data preparation: Each hospital extracted patient data from the EMR system for patients diagnosed with T2DM between 2015 and 2023, who initiated second-line treatment with SGLT-2i or DPP-4i, and had a follow-up of ≥12 months. Data volume: Hospital A: 5000 cases, Hospital B: 4800 cases, Hospital C: 4500 cases, Hospital D: 3200 cases, Hospital E: 2800 cases, totaling 20300 cases. Each hospital independently performed data cleaning, feature engineering, and cohort construction.

[0054] Scenario-related drug knowledge graph subgraph construction: Target information for SGLT-2i (dapagliflozin, empagliflozin, canagliflozin) and DPP-4i (sitagliptin, linagliptin, saxagliptin) was extracted from DrugBank. Common side effects were extracted from SIDER. Metabolic pathways were extracted from KEGG. A knowledge graph containing hypoglycemic drugs, target proteins, cardiovascular diseases, and side effects was constructed, with approximately 500 nodes and 2000 edges.

[0055] Global Drug Multi-Relationship Knowledge Graph Based on this, subgraph G′ related to second-line hypoglycemic drugs and cardiovascular safety was extracted, including: Drug pathways: SGLT-2 inhibitors (dapagliflozin, empagliflozin, canagliflozin), DPP-4 inhibitors (sitagliptin, linagliptin, saxagliptin), diuretics, ACEIs / ARBs, beta-blockers, statins, etc.; Target pathways: SGLT2 protein, cardiovascular-related receptors and ion channels, etc.; Disease pathways: T2DM, coronary artery disease, heart failure, cardiovascular events, etc.; Side effect pathways: hypotension, volume depletion, electrolyte imbalance, etc.

[0056] Simultaneously perform quality control: remove redundant and erroneous relationships, and supplement missing key cardiovascular safety-related relationships.

[0057] Graph Neural Network DDI Prediction: The R-GCN / GAT architecture is used to learn node representations on subgraph G′ to obtain drug node embeddings. Subsequently, bilinear decoders and sigmoid algorithms were used to calculate the probabilities of various interaction types for combination therapies (such as SGLT-2 inhibitors and diuretics). .

[0058] In this embodiment, the model predicts that the probability of SGLT-2i interacting with diuretics in the "volume deprivation / hypotension-related DDI" type is approximately 0.35, corresponding to an interaction probability of 35%, requiring monitoring of blood pressure and renal function. DPP-4i interacts less with other hypoglycemic agents. The system identifies the critical pathway: SGLT-2i → SGLT2 inhibition → increased urinary sodium excretion → synergistic effect with diuretics → increased risk of hypotension.

[0059] Patient-level DDI risk burden and regimen safety score: For each patient p, extract the set of concomitant medications used during the follow-up period. Calculate the DDI risk burden and solution-level comprehensive security score: ; ; in, This indicates the overall risk burden of drug freshness during the follow-up period for patient p. This indicates a candidate second-line scheme (such as the SGLT-2i scheme or the DPP-4i scheme). For the patient pool to receive this treatment, This indicates the safety score of the current medication regimen, which is part of the subsequent efficacy-safety comprehensive evaluation.

[0060] Specifically, federated learning training has multiple rounds. In the first round, each medical institution trains on its local data for 10 epochs. An epoch represents a complete training run of the model using all data in the training set, calculating the gradient for each medical institution. The gradient is encrypted using a homomorphic encryption scheme (Paillier encryption), and the encrypted local gradient is uploaded. To the central server. The central server calculates the weighted average in the encrypted domain: ,in For medical institutions The number of samples. The central server decrypts to obtain the global gradient. Update global model parameters The materials were then distributed to various hospitals. Each hospital downloaded them. Update the local model.

[0061] Repeat the above process for a total of 50 rounds of communication, with 10 epochs of local training in each round. During the training process, monitor the performance of the global validation set (10% of the data extracted from each hospital) until the model converges. Finally, the propensity score AUC of the global model reaches 0.82.

[0062] Next, distribution offset detection was performed, revealing differences in model performance across hospitals' local validation sets: Hospital A: AUC = 0.84, RMSE = 1.6; Hospital B: AUC = 0.83, RMSE = 1.7; Hospital C: AUC = 0.81, RMSE = 1.9; Hospital D: AUC = 0.79, RMSE = 2.2; Hospital E: AUC = 0.78, RMSE = 2.4. The Maximum Mean Discrepancy (MMD) test was used to assess differences in baseline covariate distributions: A-B MMD = 0.05 (small, similar distribution); A-D MMD = 0.18 (moderate); A-E MMD = 0.25 (relatively large). The reasons for this are: the patient population characteristics (e.g., ethnic composition, dietary habits, comorbidity spectrum) in the regions where Hospitals D and E are located differ from those in the regions where Hospitals A and B are located; and the standardization of hospital treatment in the regions where Hospitals D and E are located is relatively low, resulting in slightly lower data recording quality.

[0063] A causal reinforcement learning model based on federated learning was used to evaluate the efficacy and safety of SGLT-2i vs DPP-4i: Efficacy (HbA1c reduction, 3-year long-term control): SGLT-2i: HbA1c reduction of 0.8% (95% CI: 0.7-0.9); DPP-4i: HbA1c reduction of 0.6% (95% CI: 0.5-0.7); Efficacy difference: SGLT-2i was superior to DPP-4i by approximately 0.2% (P<0.001); Cardiovascular safety (3-year MACE incidence): SGLT-2i: 8.2% (95% CI: 7.5-9.0); DPP-4i: 10.5% (95% CI: 9.7-11.4); SGLT-2i significantly reduced the risk of MACE (HR=0.78, 95% CI: 0.70-0.87). P<0.001); DDI risk: SGLT-2i increases the risk of volume depletion and hypotension when used in combination with diuretics (interaction probability 35%, blood pressure and renal function need to be monitored); DPP-4i has fewer interactions with other hypoglycemic agents.

[0064] Long-term value function (overall efficacy, cardiovascular outcome, risk of hypoglycemia, adherence): SGLT-2i: ; (in months of complication-free survival); DPP-4i: Recommendation: For patients with type 2 diabetes mellitus (T2DM) without contraindications to SGLT-2 inhibitors, SGLT-2 inhibitors are superior to DPP-4 inhibitors as second-line therapy, not only in terms of better glucose lowering efficacy but also in significantly reducing the risk of cardiovascular events. However, monitoring of volume status and renal function is necessary, especially in patients using diuretics concurrently.

[0065] Federated learning enabled the integration of data from 20,300 multicenter patients while protecting the data privacy of each hospital. This significantly increased the sample size and the credibility of the evaluation conclusions, and covered populations from different regions in the east and west, resulting in stronger external validity.

[0066] Example 4

[0067] Based on Example 1, there is another application scenario for multi-center joint evaluation of drug safety. This scenario adopts the multi-center joint evaluation and system for drug safety based on graph neural networks and federated learning as described in the above examples. This example demonstrates how a provincial drug regulatory agency uses the system of this invention to evaluate the real-world efficacy and bleeding risk of a novel oral anticoagulant (NOAC).

[0068] Specifically, this NOAC has been approved for stroke prevention in patients with nonvalvular atrial fibrillation. In the three years since its market launch, it has been used by approximately 100,000 patients. The regulatory agency has initiated a post-marketing evaluation to compare the efficacy (stroke prevention) and safety (bleeding risk) of the NOAC versus conventional warfarin.

[0069] First, data were collected from 50 hospitals from 2020 to 2023 on patients with non-valvular atrial fibrillation who had started NOAC or warfarin treatment and were followed up for ≥6 months, including a total of 35,000 patients (22,000 in the NOAC group and 13,000 in the warfarin group).

[0070] Next, the inclusion and exclusion criteria, outcomes, and covariate definitions are determined, specifically as follows: Inclusion criteria: Adults with nonvalvular atrial fibrillation, first initiation of NOAC or warfarin, and records of key covariates and outcomes; Exclusion criteria: valvular atrial fibrillation, severe liver failure, pregnancy, contraindications to anticoagulation, etc. Effective outcome: ischemic stroke or systemic embolism; Safety outcomes: major bleeding (as defined by the ISTH Hemostasis Association), intracranial hemorrhage, gastrointestinal bleeding, etc. Covariate: CH D -VASc score, HAS-BLED score, comorbidities, and use of antiplatelet drugs, proton pump inhibitors, amiodarone, etc.

[0071] Next, the knowledge graph related to NOACs and the probability of DDI (Drawback Difference) are calculated: First, subgraphs related to NOACs, warfarin, and key concomitant drugs are extracted from the global drug knowledge graph, including drug nodes (NOACs, warfarin, aspirin, clopidogrel, proton pump inhibitors, amiodarone, etc.), target nodes (coagulation factor Xa, vitamin K circulation-related enzymes, etc.), metabolic enzyme nodes (CYP3A4, P-gp), and bleeding-related side effect nodes. The probabilities of NOACs and their concomitant drugs on the "bleeding risk-related DDI" are calculated using R-GCN or GAT with a bilinear decoder. The calculation results are as follows: The probability of a major bleeding-related DDI when NOAC is used in combination with antiplatelet drugs is approximately 0.42, which corresponds to a significantly increased risk of bleeding. The probability of DDI when NOAC is used in combination with proton pump inhibitors is close to the background level, indicating that there is no significant pharmacokinetic interaction. When NOAC is used in combination with amiodarone, it is predicted that the NOAC blood concentration will increase by about 30%, supporting the clinical dose reduction recommendation.

[0072] Next, key interaction pathways were identified and their mechanisms explained: Based on graph attention mechanisms, high-contribution pathways for the combined use of NOAC and antiplatelet drugs were identified, specifically: NOAC → inhibition of coagulation factor Xa → weakening of coagulation cascade → inhibition of platelet aggregation by antiplatelet drugs → synergistic pharmacodynamic effects → increased bleeding risk.

[0073] Based on this approach, the comparative effectiveness of efficacy and bleeding risk was analyzed: first, an active-controlled new user design was used to construct a propensity score model. ,in, This represents the patient's baseline covariate vector. Indicates a tendency rating. Indicates covariate conditions The probability of the current patient using NOAC is used to train unified model parameters through federated learning; weights are assigned according to inverse probability. Where w represents the inverse probability weighting weight, This represents the baseline covariate vector for patient i. The propensity score for patient i represents the probability of using NOAC, and q represents the actual usage status. When q=0, NOAC is recommended, and when q=1, warfarin is recommended. A weighted Cox model is built to estimate the hazard ratio (HR) and 95% CI (confidence interval) for outcomes such as ischemic stroke, major hemorrhage, and intracranial hemorrhage.

[0074] Based on the above analysis, the final evaluation results are as follows: Ischemic stroke: NOAC vs warfarin, HR≈0.78; Major hemorrhage: HR≈0.60; Intracranial hemorrhage: HR≈0.33; Annual incidence rate example: Ischemic stroke 1.8% vs 2.3%, Major hemorrhage 2.1% vs 3.5%, Intracranial hemorrhage 0.4% vs 1.2%.

[0075] By employing a propensity score model and an inverse probability weighting method, baseline characteristics of different treatment groups were balanced. Then, a weighted Cox model was used to estimate the hazard ratios of various clinical outcomes, enabling a comparative efficacy and safety analysis of two candidate treatment regimens in a real-world setting. Results showed that NOACs were superior to warfarin in both efficacy and bleeding risk, demonstrating the application value of this invention in the joint evaluation of efficacy and safety in real-world settings.

[0076] Based on the above evaluation results, a real-world evidence report is generated. Specifically, the system automatically generates a standardized RWE report, including sections on data sources, research design, bias control, statistical methods, main results, and DDI risk warnings. Data source auditing records the complete process from raw data extraction to final conclusions, ensuring reproducibility.

[0077] Example 5

[0078] like Figure 3 As shown in Example 1, Example 3 proposes a terminal device for a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning. The terminal device 200 includes at least one memory 210, at least one processor 220, and a bus 230 connecting different platform systems.

[0079] The memory 210 may include a readable medium in the form of volatile memory, such as RAM 211 and / or cache memory 212, and may further include ROM 213.

[0080] The memory 210 also stores a computer program that can be executed by the processor 220, causing the processor 220 to execute any of the above-described methods for multi-center joint evaluation of drug safety based on graph neural networks and federated learning in this application embodiment. The specific implementation and technical effects are consistent with those described in the embodiments of the above methods, and some details will not be repeated here. The memory 210 may also include a program / utility 214 having a set (at least one) of program modules 215. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0081] Accordingly, processor 220 can execute the aforementioned computer program, as well as executable program / utility 214.

[0082] Bus 230 can represent one or more of several types of bus structures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or a local bus using any of the various bus structures.

[0083] Terminal device 200 can also communicate with one or more external devices 240, such as keyboards, pointing devices, Bluetooth devices, etc., and with one or more devices capable of interacting with it, and / or with any device that enables it to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed via I / O interface 250. Furthermore, terminal device 200 can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 260. Network adapter 260 can communicate with other modules of terminal device 200 via bus 230. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal device 200, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0084] Example 6

[0085] like Figure 4 As shown in Example 1, this example proposes a computer-readable storage medium for a multi-center joint evaluation method for drug safety based on graph neural networks and federated learning. The computer-readable storage medium stores instructions that, when executed by a processor, implement any of the aforementioned multi-center joint evaluation methods for drug safety based on graph neural networks and federated learning. The specific implementation method and the achieved technical effects are consistent with those described in the examples above, and some details will not be repeated.

[0086] The program product 300 provided in this embodiment for implementing the above method can be a portable compact disc read-only memory (CD-ROM) and include program code, and can run on a terminal device, such as a personal computer. However, the program product 300 of the present invention is not limited thereto. In this embodiment, the readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The program product 300 can employ any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0087] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0088] This invention is described from the perspectives of its intended use, effectiveness, progress, and novelty. Its practical and progressive features meet the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

[0089] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-center joint evaluation of drug safety based on a graph neural network and federated learning, characterized in that, The method comprises the following steps: S1. Constructing a drug multi-relationship knowledge graph based on a drug multi-source public database; S2. Constructing a graph neural network model and performing representation learning according to the drug multi-relationship knowledge graph to obtain embedding vectors of nodes; S3. Predicting the interaction between drugs according to the embedding vectors of the nodes to obtain the probability of the interaction between drugs and the type of the interaction between drugs; S4. Calculating the safety score of the candidate drug scheme according to the corresponding weight of the type of the interaction between drugs and the probability of the interaction between drugs; S5. Calculating the contribution of each edge of the graph neural network to the prediction result of the drug interaction by a graph attention algorithm according to the safety score, and generating an interpretable drug interaction mechanism description by combining a natural language description algorithm; S6. Training the graph neural network model by performing steps S1-S5 based on local real-world data in multiple medical centers to calculate the model parameter gradient of each medical center, uploading and aggregating the encrypted model parameter gradient of each medical center through a federated learning framework to obtain the global graph neural network model parameters of multiple centers, and updating the graph neural network model of each medical center until the model converges; S7. Calculating the distribution deviation difference value of the interpretable drug interaction mechanism description of each medical center by a covariate distribution test algorithm according to the global model parameters of multiple centers, and when the distribution deviation difference value exceeds a preset threshold, prompting that the interpretable drug interaction mechanism description of the current medical center has uncertainty for the current patient, generating a drug safety evaluation report and giving a warning.

2. The method according to claim 1, wherein, Step S1 comprises the following sub-steps: S101. Extracting drug relationship data from the drug multi-source public database, wherein the drug relationship data comprises drug target relationship data, drug basic information data, drug side effect relationship data, drug pathway relationship data, drug-drug interaction relationship data, and protein-protein relationship data; S102. After standardizing the extracted drug relationship data, mapping the drug entities to a unified drug coding system, and performing entity disambiguation and fusion on different representations of the same drug entity; S103. Constructing a heterogeneous graph representation according to the drug coding system after entity disambiguation and fusion to obtain a drug multi-relationship knowledge graph, which is represented as: ； wherein V represents a node set, E represents an edge set, and R represents a relationship type set.

3. The method according to claim 1, wherein, Step S2 comprises the following sub-steps: S201: According to the drug multi-relationship knowledge graph, modeling the embedding learning of the drug multi-relationship knowledge graph by a relational graph convolution network; S202: Based on a message passing mechanism, aggregating the neighbor information of the nodes in the relational graph convolution network according to the drug relationship type, and updating the embedding of the nodes, wherein the node embedding update formula is represented as: ； wherein, representing a node In the first layer embedding vectors, representing a drug relation type, representing a set of drug relation types, representing a drug relation and a neighbor node set connected with the node , and respectively represent a learnable weight matrix, representing an activation function; S203: Repeating step S202, after multiple layers of message passing, the embedding vector of each node fuses the structure and feature information in the multi-hop neighborhood to obtain the embedding vector of each node.

4. The method according to claim 1, wherein, The prediction of the drug-drug interaction in step S3 includes: by a bilinear decoder, interaction calculation is performed on the node embedding vectors of any two drugs i and j in the candidate drug use scheme to predict the occurrence probability of the drug-drug interaction, and the interaction calculation formula is represented as: ； wherein, denotes the interaction occurrence probability, and denote the drug nodes, denotes the interaction relation type, and denote the node embedding vectors, T denotes the matrix transpose, denotes the interaction relation specific weight matrix, the interaction occurrence probability is normalized by the sigmoid function, and the formula is represented as: ； ； where DDI represents drug-drug interaction, represents the probability of drug i and drug j interacting in the interaction type on the upper side, represents the sigmoid function, represents the drug pair, represents the total number of drugs, and the interaction prediction is performed for all drug pairs in the candidate drug regimen to obtain the probability of drug i and drug j interacting in each drug-drug interaction type.

5. The method according to claim 1, wherein, The calculation in step S4 to obtain the safety score of the candidate drug regimen includes assigning a severity weight to each interaction type The safety score calculation formula is represented as: ； wherein, denotes the safety score, denotes the set of drug relationship types, DDI denotes drug-drug interactions, denotes the probability value that drug i and drug j interact in the interaction relationship type r, is the severity weight of the interaction type r, determined according to a pre-set clinical severity grading.

6. The method according to claim 1, wherein the method is characterized in that, Step S5 includes the following sub-steps: S501: According to the safety score, the attention weight of each edge of the graph neural network is calculated through the graph attention mechanism, the maximum attention weight is obtained, and the corresponding edge is identified as the propagation path with the largest contribution between drugs and drugs; S502: According to the propagation path obtained in step S501, the intermediate nodes and drug relationship types on the path are extracted, and an interpretable drug interaction mechanism description is generated based on the propagation path information through a natural language description algorithm.

7. The method according to claim 1, wherein the method is characterized in that, Step S6 includes the following sub-steps: S601: Deploy the graph neural network model in the local training environment of multiple medical centers, and based on the local real world data, repeat steps S1-S5 to train the model, and obtain the gradient update of the graph neural network model parameters of each medical center; S602: Based on the pre-set secure multi-party computation protocol, each medical center encrypts the graph neural network model parameter gradient update and transmits it to the central aggregation server; S603: According to the encrypted graph neural network model parameter gradient update uploaded by each medical center, the central aggregation server performs federated average aggregation to generate a global graph neural network model parameter of multiple centers; S604: According to the global model parameter, update the local model of each medical center; S605: Repeat steps S602 to S604 until the model converges.

8. The method according to claim 1, wherein the method is characterized in that, Step S7 includes the following sub-steps: S701: According to the multi-center global model parameter, the distribution of the patient population characteristics of each medical center is quantitatively analyzed through the covariate distribution test algorithm, and the distribution difference of the patient population characteristics between different medical centers is obtained; S702: According to the distribution difference of the patient population characteristics between different medical centers, compare the model performance indicators of the local real world data of each medical center to obtain the distribution offset difference value of the interpretable drug interaction mechanism description of each medical center; S703: If the distribution offset difference value exceeds the preset threshold value, it is prompted that the interpretable drug interaction mechanism description of the current medical center has uncertainty for the current patient, a drug safety evaluation report is generated and a warning is given.

9. The method according to claim 1, wherein, The drug safety evaluation report in step S7 includes data source description, bias control method, missing data processing strategy, and queue construction process, data traceability audit clues; the data source description includes medical center number, data time range, and patient population characteristics; The bias control method includes selective bias control method and confounding bias control method; the data traceability audit clues include the complete processing flow record from step S1 to step S7.

10. A drug safety multi-center joint evaluation system based on a graph neural network and federated learning, the system being implemented based on the drug safety multi-center joint evaluation method based on a graph neural network and federated learning according to any one of claims 1-9, characterized in that, It includes: A knowledge graph construction module for constructing a drug multi-relation knowledge graph based on a drug multi-source public database; A graph neural network construction module for constructing a graph neural network model based on the drug multi-relation knowledge graph and performing representation learning to obtain the embedding vectors of each node; An interaction prediction module is configured to predict the drug-drug interactions based on the embedding vectors of the nodes, to obtain a drug-drug interaction occurrence probability and a drug-drug interaction type; A safety evaluation module is configured to calculate a safety score of the candidate drug regimen based on the corresponding weight of the drug-drug interaction type and the drug-drug interaction occurrence probability; An interpretable mechanism generation module is configured to calculate, based on the safety score, a contribution degree of each edge of the graph neural network to the drug-drug interaction prediction result by using a graph attention algorithm, and to generate an interpretable drug-drug interaction mechanism description by using a natural language description algorithm; A federated learning coordination module is configured to train the graph neural network model based on the local real-world data of multiple medical centers by performing steps S1-S5, to calculate model parameter gradients of the medical centers, to encrypt and upload and aggregate the model parameter gradients of the medical centers by using a federated learning framework, to obtain global graph neural network model parameters of the multiple medical centers, and to update the graph neural network models of the medical centers until the models converge; A distribution shift detection and warning module is configured to calculate, based on the global model parameters of the multiple medical centers, a distribution shift difference value of the interpretable drug-drug interaction mechanism description of each medical center by using a covariate distribution test algorithm, to prompt that the interpretable drug-drug interaction mechanism description of the current medical center has uncertainty for the current patient when the distribution shift difference value exceeds a preset threshold, to generate a drug safety evaluation report, and to perform a warning.

Images