Patient-level adverse drug reaction prediction method and system based on graph neural network

Abstract

The present invention relates to the technical field of drug adverse reaction prediction, and discloses a patient-level drug adverse reaction prediction method and system based on a graph neural network. An adverse reaction prediction model is used for prediction. When the adverse reaction prediction model is trained, sample data is input into a patient feature encoder for feature learning to obtain basic features; the sample data is used to obtain embedded representations of entities and relationships through a knowledge graph, and the embedded representation is used as the input of a graph neural network model for feature learning to obtain enhanced features; based on the basic features and enhanced features, adverse drug reactions are predicted through a prediction model. The association between patients and diseases, drugs, and adverse reactions is constructed by introducing a knowledge graph, and the knowledge graph is analyzed using a graph neural network to better predict the adverse reactions of patients. Accurate prediction of adverse drug reactions based on the individual patient's own conditions is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of adverse drug reaction prediction, and in particular to a patient-level adverse drug reaction prediction method and system based on a graph neural network. Background Art

[0002] Currently, precision medicine has become an important development direction. Precision medicine aims to provide each patient with personalized diagnosis, prevention and treatment plans based on individual patient characteristics such as genes, biomarkers and lifestyle. However, to achieve precision medicine, many challenges must be overcome, including accurately predicting patients' adverse reactions to drugs.

[0003] Adverse Drug Reaction (ADR) is one of the serious problems facing the medical field. It not only increases the pain and treatment costs of patients, but may also lead to serious health problems or even death. Therefore, one of the keys to achieving precision medicine is to improve the accuracy of predicting patients' adverse drug reactions.

[0004] In the field of adverse drug reaction (ADR) prediction, the current focus is on using machine learning methods for prediction. These methods usually use data such as the chemical structure, genetic information, and protein interactions of drugs as features to build models to predict possible side effects of drugs. However, existing methods have limitations and can only provide general prediction results, without accurate predictions for individual patients' own conditions.

[0005] Therefore, there is an urgent need for patient-level adverse drug reaction prediction methods and systems based on graph neural networks to achieve patient-level adverse drug reaction prediction and improve the accuracy of adverse drug reaction prediction. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a patient-level adverse drug reaction prediction method based on graph neural network, comprising the following steps:

[0007] Acquire a historical data set as sample data for training an adverse drug reaction predictor; the sample data includes multiple groups of patient information and corresponding adverse drug reactions;

[0008] Using the sample data to train the adverse drug reaction predictor is specifically as follows:

[0009] Inputting the sample data into a patient feature encoder for feature learning to obtain basic features;

[0010] Obtaining embedded representations of entities and relationships from the sample data through a knowledge graph; the entities include patient nodes, disease nodes, and drug nodes, and the relationships include therapeutic relationships between diseases and drugs, side effect relationships between drugs and adverse drug reactions, and association relationships between patients and diseases and drugs, and association relationships between patients and adverse drug reactions;

[0011] Using the embedding representation as an input of a graph neural network model to perform feature learning to obtain enhanced features;

[0012] Predicting adverse drug reactions through a prediction model according to the basic features and the enhanced features until the loss function reaches a preset convergence speed, thereby completing the training of the adverse drug reaction predictor;

[0013] The patient information is input into the trained adverse drug reaction predictor to perform patient-level adverse drug reaction prediction.

[0014] Furthermore, the patient feature encoder is a neural network model.

[0015] Further, the historical data set is stored in the form of groups, each group of data includes patient information of a corresponding patient and a corresponding adverse drug reaction;

[0016] The patient information includes a set of diseases suffered by the corresponding patient, a set of medications taken, age, gender, and weight.

[0017] Furthermore, the prediction model is used to predict adverse drug reactions, and the calculation formula is:

[0018] ;

[0019] In the formula, represents the set of adverse drug reaction probabilities for each patient predicted by the prediction model, Represents enhanced features, Represents the basic features, Represents a predictor.

[0020] Furthermore, when training the adverse drug reaction predictor, the loss function L is calculated as follows:

[0021] ;

[0022] In the formula, α represents the first hyperparameter, represents the loss function of the graph neural network, Represents the loss function of the patient feature encoder.

[0023] Furthermore, the loss function of the graph neural network is expressed using InfoNCE, and the calculation formula is as follows:

[0024] ;

[0025] In the formula, B represents the number of data loaded in each batch, sim represents the cosine similarity calculation, and exp() represents the exponential function. represents the enhanced features of the i-th data loaded in the current batch, Represents the basic features of the i-th data when the batch is loaded, Represents the basic features of the j-th data loaded in the batch.

[0026] Furthermore, the loss function of the patient feature encoder is expressed using Focal, and the calculation formula is as follows:

[0027] ;

[0028] Where γ is the second hyperparameter, represents the probability set of adverse reactions in historical data, Represents the probability set of adverse reactions output by the prediction model.

[0029] The present application also provides a patient-level adverse drug reaction prediction system based on a graph neural network, which executes any of the above-mentioned patient-level adverse drug reaction prediction methods based on a graph neural network, and the system includes:

[0030] A basic feature learning module, used for performing feature learning through a patient feature encoder according to patient information to obtain basic features;

[0031] An enhanced feature learning module is used to obtain embedded representations of entities and relationships through a knowledge graph based on patient information, and use the embedded representations as inputs to a graph neural network model to perform feature learning to obtain enhanced features;

[0032] The adverse drug reaction prediction module is used to predict adverse drug reactions at the patient level through a prediction model based on the basic features and the enhanced features.

[0033] The embodiments of the present invention have the following technical effects:

[0034] In the present application, the patient-level adverse drug reaction prediction method based on graph neural network uses the adverse drug reaction predictor for prediction. When training the adverse drug reaction predictor, the sample data is input into the patient feature encoder for feature learning to obtain basic features; the sample data is obtained through the knowledge graph to obtain the embedded representation of entities and relationships, and the embedded representation is used as the input of the graph neural network model for feature learning to obtain enhanced features; according to the basic features and enhanced features, the adverse drug reaction is obtained through the prediction model. The entities include patient nodes, disease nodes, and drug nodes, and the relationships include the therapeutic relationship between disease and drug, the side effect relationship between drug and adverse reaction, and the association relationship between patient and disease and drug. The association relationship between patient and disease, drug, and adverse reaction is constructed by introducing the knowledge graph, and the graph neural network is used to analyze the knowledge graph to better predict the adverse reaction of the patient. The patient information including the set of diseases suffered by the corresponding patient, the set of drugs taken, age, gender, and weight is input into the trained adverse drug reaction predictor, and the patient-level adverse drug reaction prediction can be completed, and the individual patient's own situation is accurately predicted, which further improves the prediction accuracy of adverse drug reactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 is a flowchart of a method for predicting adverse drug reactions at the patient level based on a graph neural network provided by an embodiment of the present invention;

[0037] Figure 2 is a flow chart of a method for training an adverse reaction prediction model provided by an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of the framework of the adverse drug reaction predictor provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. 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 belong to the scope of protection of the present invention.

[0040] When using machine learning methods to predict adverse drug reactions (ADRs), data such as the drug's chemical structure, genetic information, and protein interactions are usually used as features to build models to predict possible side effects of drugs. However, since the individual patient's own situation is not taken into account, the accuracy of the prediction results is reduced. This application introduces a knowledge graph to construct the relationship between patients and diseases, drugs, and adverse reactions, and uses a graph neural network to analyze the knowledge graph to better predict patients' adverse reactions.

[0041] Depend on Figure 1 and Figure 3 It can be seen that the patient-level adverse drug reaction prediction method based on graph neural network includes the following steps:

[0042] S1: Obtain a historical data set as sample data for training an adverse drug reaction predictor; the sample data includes multiple groups of patient information and corresponding adverse drug reactions;

[0043] Further, the historical data set is stored in the form of groups, each group of data includes patient information of a corresponding patient and a corresponding adverse drug reaction;

[0044] The patient information includes a set of diseases suffered by the corresponding patient, a set of medications taken, age, gender, and weight. Combined with the patient's disease, medication, age, gender, weight and other information, the adverse reactions that different factors may cause to the patient are comprehensively considered.

[0045] Exemplarily, the historical data set includes N groups of data, wherein the nth group of data is the patient information of a patient n. And the adverse reaction set of this patient .in Is a patient A collection of diseases, Is a patient Collection of medications, Is a patient Age, Is a patient Gender (where 1 indicates biological sex is male and 2 indicates biological sex is female), Is a patient weight.

[0046] S2: Use the sample data to train the drug adverse reaction predictor, such as Figure 2 As shown, specifically:

[0047] S21: inputting the sample data into a patient feature encoder for feature learning to obtain basic features;

[0048] Furthermore, the patient feature encoder is a neural network model.

[0049] S22: Obtaining embedding representations of entities and relationships from the sample data through a knowledge graph; the entities include patient nodes, disease nodes, and drug nodes; the relationships include therapeutic relationships between diseases and drugs, side effect relationships between drugs and adverse drug reactions, and association relationships between patients and diseases and drugs, and association relationships between patients and adverse drug reactions;

[0050] A knowledge graph is a structured semantic knowledge base that represents the relationship between entities in a graphical manner. This application introduces a knowledge graph for modeling and utilizing the relationships between patients, diseases, drugs, and ADRs. As a basic component, the knowledge graph helps to accurately predict by capturing the complex associations between patients, diseases, and drugs. In the knowledge graph, individual nodes are used to represent patients, diseases, and drugs, while edges represent the connection between patients who report a disease and patients who are prescribed a specific drug. Specifically, this application uses bag-of-words (BOW) features to represent the node features of patients. In contrast, the node features of the disease and the node features of the drug are described using unique hot features.

[0051] S23: Using the embedded representation as an input of the graph neural network model to perform feature learning to obtain enhanced features;

[0052] Graph neural network is a machine learning model specifically designed to process graph structured data. It can learn the representation of nodes, edges, and the entire graph in the graph, and is used for tasks such as node classification, graph classification, and link prediction. Graph neural networks can be used to learn the embedded representation of entities and relationships in knowledge graphs, and then used for tasks such as entity linking, relationship prediction, or graph completion of knowledge graphs. This combination utilizes the structured information of knowledge graphs and the learning ability of graph neural networks to improve the level of intelligent application of knowledge graphs. By using graph neural networks, accurate prediction of adverse drug reactions can be achieved while taking into account individual patient characteristics, thereby providing better services for precision medicine.

[0053] S24: predicting adverse drug reactions through a prediction model according to the basic features and the enhanced features until the loss function reaches a preset convergence speed, thereby completing the training of the adverse drug reaction predictor;

[0054] The preset convergence speed can be expressed as the loss function remains unchanged within the preset number of steps. The preset number of steps can be set according to needs.

[0055] Exemplarily, the adverse drug reaction predicted by the prediction model may be a probability set of adverse drug reactions, that is, a set of probabilities of various adverse drug reactions that may be caused.

[0056] Furthermore, the prediction model is used to predict adverse drug reactions, and the calculation formula is:

[0057] ;

[0058] In the formula, represents the set of adverse drug reaction probabilities for each patient predicted by the prediction model, Represents enhanced features, Represents the basic features, Represents a predictor.

[0059] represents the adverse drug reactions output by the prediction model, Represents enhanced features, Indicates basic features.

[0060] The adverse drug reaction predictor enables the effective aggregation of heterogeneous patient data, ultimately improving the accuracy of ADR prediction. Heterogeneous patient data means that each node can be of different types, such as a patient is a node and a disease is a node.

[0061] Furthermore, when training the adverse drug reaction predictor, the loss function L is calculated as follows:

[0062] ;

[0063] In the formula, α represents the first hyperparameter, represents the loss function of the graph neural network, Represents the loss function of the patient feature encoder.

[0064] Furthermore, the loss function of the graph neural network is expressed using InfoNCE, and the calculation formula is as follows:

[0065] ;

[0066] In the formula, B represents the number of data loaded in each batch, sim represents the cosine similarity calculation, and exp() represents the exponential function. represents the enhanced features of the i-th data loaded in the current batch, Represents the basic features of the i-th data when the batch is loaded, Represents the basic features of the j-th data loaded in the batch.

[0067] InfoNCE (Information Noise Contrastive Estimation) is a loss function for self-supervised learning, which plays an important role in unsupervised or self-supervised representation learning. The core idea of ​​this loss function is to learn useful feature representations by maximizing the similarity between the target sample and the positive sample, while minimizing the similarity between the target sample and a set of negative samples.

[0068] InfoNCE is applied to maximize the consistency between the representations generated by the GNN graph neural network and the patient feature encoder, and minimize the consistency between the representations of irrelevant data samples. The InfoNCE loss is shown in the above formula.

[0069] Furthermore, the loss function of the patient feature encoder is expressed using Focal, and the calculation formula is as follows:

[0070] ;

[0071] Where γ is the second hyperparameter, represents the probability set of adverse reactions in historical data, Represents the probability set of adverse reactions output by the prediction model.

[0072] Focal refers to the specific loss function Focal Loss, which can be used to deal with the loss function of class imbalance. The patient-level ADR prediction task is regarded as a multi-class multi-label classification problem. This application uses Focal Loss as the training target for dealing with unbalanced ADR.

[0073] For example, Figure 3 As shown, the adverse drug reaction predictor consists of several key components, including a graph aggregation layer, a patient node enhancement layer, and a predictor.

[0074] In the context of knowledge graphs, node features belong to different types and domains. First, these features are projected from their respective domains to a shared domain. The conversion process is as follows:

[0075] ;

[0076] LayerNorm represents regular calculation. Indicates the node type, P indicates the patient node, D indicates the disease node, and M indicates the drug node. represents the characteristics of the initial layer, Indicates Characteristics of the layer.

[0077] Next, a graph convolution layer is used to promote message aggregation in the knowledge graph. This application uses HGT to build the default graph convolution layer. HGT (Heterogeneous Graph Transformer) is a graph neural network model for processing heterogeneous graph data. It is capable of processing graphs containing multiple types of nodes and edges, which can represent different entities and the relationships between them. HGT can effectively capture and model the heterogeneity in the graph by introducing heterogeneous mutual attention mechanism and heterogeneous message passing mechanism. In addition, HGT can also process dynamic graph data and capture the temporal dependencies between nodes and edges in the graph through relative time encoding technology.

[0078] The process of each layer of HGT can be broken down into the following core components:

[0079] ;

[0080] Where s represents the source node, t represents the target node, represents the first-order neighbors of the target node, Represents the set of all edges from source node s to target node t. There are three basic operations: , used to estimate the importance of each source node s; Used to extract messages by using source node s; Used to aggregate all source nodes by attention weights Neighborhood news; Represents the representation of the target node t after HGT processing.

[0081] This application first calculates the attention, expressed as , which represents the importance of each source node s relative to the target node t. This mechanism uses the Query-Value framework inspired by the Transformer architecture. The formula for attention calculation is as follows:

[0082] ; ;

[0083] ;

[0084] ;

[0085] here Indicates the node type of the target node t; is a trainable vector representing each meta-relation triple the general importance of Represents a trainable weight matrix. Softmax is the softmax function; Here, the information of h heads is concatenated. The multi-head attention mechanism is used here. In particular, in the mth attention head In the context of The source node s of type is converted into the mth vector is done by applying a linear projection, expressed as The superscript m indicates the mth Attention head, and the subscript Indicates the node type. Specifically, for the same Attention Head model content, different linear projections will be used for different node types. This linear projection will map the original d-dimensional features to d / h dimensions for dimensionality reduction. After dimensionality reduction, h heads are concatenated to obtain d-dimensional features again. Here, h represents the number of attention heads, and d / h corresponds to the vector dimensions of each attention head. It is important to note that The index of is based on the source node type τ(s). This means that each type of node has a unique linear projection. This unique feature enables the model to adjust and tailor its projection to the specific type of node, maximizing its ability to capture differences in the internal distribution of the data. This approach helps the model effectively represent and understand the heterogeneous information between different node types.

[0086] Once the attention is determined, we proceed to extract the message, marked as , based on the attention weights. The message is calculated by linear projection, taking into account the source node s and the dependencies of the edges. Linear projection is also used Project the target node t into the mth query vector. The process is described in detail as follows:

[0087] ;

[0088] ;

[0089] here is the edge type of the edge in the knowledge graph. Specifically, to obtain the mth message header , first by using the linear transformation Project the source node s of type τ(s) into the mth message vector. Next, introduce a matrix , integrating edge dependencies into message representation.

[0090] The last step is to concatenate all h message headers to form the Through a series of After the heterogeneous graph aggregation layer completes this process, the enhanced feature representation is obtained, which is expressed as .

[0091] Subsequently, a fully connected layer, called the patient node augmentation network, is integrated into This enhancement process introduces random noise to the original features during model training, defined as Among them, X is the original feature, E is the all-1 matrix, and the Dropout() function will target each element with probability The original value is set to 0, and the whole formula is to add a certain amount of noise to the original feature X. However, in the prediction using the adverse reaction prediction model, the dropout calculation is not applied to prevent overfitting. Finally, a fully connected layer is used to predict the ADR of each patient.

[0092] S3: Input the patient information into the trained adverse drug reaction predictor to perform patient-level adverse drug reaction prediction.

[0093] The present application also provides a patient-level adverse drug reaction prediction system based on a graph neural network, which executes any of the above-mentioned patient-level adverse drug reaction prediction methods based on a graph neural network, and the system includes:

[0094] A basic feature learning module, used for performing feature learning through a patient feature encoder according to patient information to obtain basic features;

[0095] An enhanced feature learning module is used to obtain embedded representations of entities and relationships through a knowledge graph based on patient information, and use the embedded representations as inputs to a graph neural network model to perform feature learning to obtain enhanced features;

[0096] The adverse drug reaction prediction module is used to predict adverse drug reactions at the patient level through a prediction model based on the basic features and the enhanced features.

[0097] In the present application, the patient-level adverse drug reaction prediction method based on graph neural network uses the adverse drug reaction predictor for prediction. When the adverse drug reaction predictor is trained, the sample data is input into the patient feature encoder for feature learning to obtain basic features; the sample data is obtained through the knowledge graph to obtain the embedded representation of entities and relationships, and the embedded representation is used as the input of the graph neural network model for feature learning to obtain enhanced features; according to the basic features and enhanced features, the adverse drug reaction is obtained through the prediction model. The entities include patient nodes, disease nodes, and drug nodes, and the relationships include the therapeutic relationship between disease and drug, the side effect relationship between drug and adverse reaction, and the association relationship between patient and disease and drug. The association relationship between patient and disease, drug, and adverse reaction is constructed by introducing the knowledge graph, and the graph neural network is used to analyze the knowledge graph to better predict the adverse reaction of the patient. The patient information including the set of diseases suffered by the corresponding patient, the set of drugs taken, age, gender, and weight is input into the trained adverse drug reaction predictor, and the patient-level adverse drug reaction prediction can be completed, and the individual patient's own situation is accurately predicted, which further improves the prediction accuracy of adverse drug reactions.

[0098] This application achieves more accurate patient-level prediction of adverse drug reactions, which can help medical institutions better identify and manage patients' adverse reaction risks and improve the level of medical services. Early prediction and prevention of adverse reactions that may occur in patients can reduce unnecessary medical expenses and treatment time, and reduce the waste of medical resources. The successful application of the solution of this application will help promote the development of precision medicine and provide technical support for individualized treatment.

[0099] It should be noted that the terms used in the present invention are only for describing specific embodiments, rather than limiting the scope of the present application. As shown in the present specification, unless the context clearly indicates an exception, the words "one", "a", "a kind of" and / or "the" do not specifically refer to the singular, but may also include the plural. The terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method or device. In the absence of more restrictions, the elements defined by the sentence "include one..." do not exclude the presence of other identical elements in the process, method or device including the elements.

[0100] It should also be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", etc. should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0101] 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 by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A patient-level adverse drug reaction prediction method based on graph neural network, characterized in that: The steps include: Acquire a historical data set as sample data for training an adverse drug reaction predictor; the sample data includes multiple groups of patient information and corresponding adverse drug reactions; Using the sample data to train the adverse drug reaction predictor is specifically as follows: Inputting the sample data into a patient feature encoder for feature learning to obtain basic features; Obtaining embedded representations of entities and relationships from the sample data through a knowledge graph; the entities include patient nodes, disease nodes, and drug nodes, and the relationships include therapeutic relationships between diseases and drugs, side effect relationships between drugs and adverse drug reactions, and association relationships between patients and diseases and drugs, and association relationships between patients and adverse drug reactions; Using the embedding representation as an input of a graph neural network model to perform feature learning to obtain enhanced features; Predicting adverse drug reactions through a prediction model according to the basic features and the enhanced features until the loss function reaches a preset convergence speed, thereby completing the training of the adverse drug reaction predictor; The patient information is input into the trained adverse drug reaction predictor to perform patient-level adverse drug reaction prediction.

2. The method for predicting adverse drug reactions at the patient level based on graph neural network according to claim 1, characterized in that: The patient feature encoder is a neural network model.

3. The patient-level adverse drug reaction prediction method based on graph neural network according to claim 1, characterized in that: The historical data set is stored in the form of groups, each group of data includes patient information of the corresponding patient and the corresponding adverse drug reaction; the patient information includes a set of diseases suffered by the corresponding patient, a set of drugs taken, age, gender, and weight.

4. The patient-level adverse drug reaction prediction method based on graph neural network according to claim 1, characterized in that: The adverse drug reactions are predicted by the prediction model, and the calculation formula is: ； In the formula, represents the set of adverse drug reaction probabilities for each patient predicted by the prediction model, Represents enhanced features, Represents the basic features, Represents a predictor.

5. The patient-level adverse drug reaction prediction method based on graph neural network according to claim 1, characterized in that: When training the adverse drug reaction predictor, the loss function L is calculated as: ； In the formula, α represents the first hyperparameter, represents the loss function of the graph neural network, Represents the loss function of the patient feature encoder.

6. The method for predicting adverse drug reactions at the patient level based on graph neural network according to claim 5, characterized in that: The loss function of the graph neural network is expressed using InfoNCE, and the calculation formula is as follows: ； In the formula, B represents the number of data loaded in each batch, sim represents the cosine similarity calculation, and exp() represents the exponential function. represents the enhanced features of the i-th data loaded in the current batch, Represents the basic features of the i-th data when the batch is loaded, Represents the basic features of the j-th data loaded in the batch.

7. The method for predicting adverse drug reactions at the patient level based on graph neural network according to claim 5, characterized in that: The loss function of the patient feature encoder is expressed using Focal, and the calculation formula is as follows: ; Where γ is the second hyperparameter, represents the probability set of adverse reactions in historical data, Represents the probability set of adverse reactions output by the prediction model.

8. A patient-level adverse drug reaction prediction system based on a graph neural network, used to execute the patient-level adverse drug reaction prediction method based on a graph neural network according to any one of claims 1 to 7, characterized in that: The system comprises: A basic feature learning module, used for performing feature learning through a patient feature encoder according to patient information to obtain basic features; An enhanced feature learning module is used to obtain embedded representations of entities and relationships through a knowledge graph based on patient information, wherein the entities include patient nodes, disease nodes, and drug nodes, and the relationships include therapeutic relationships between diseases and drugs, side effect relationships between drugs and adverse drug reactions, and association relationships between patients and diseases and drugs, and association relationships between patients and adverse drug reactions; the embedded representation is used as an input of a graph neural network model to perform feature learning to obtain enhanced features; The adverse drug reaction prediction module is used to predict adverse drug reactions at the patient level through a prediction model based on the basic features and the enhanced features.

Images