Adverse drug reaction prediction method and system with delayed feedback

Abstract

The invention relates to an adverse drug reaction prediction method and system with delayed feedback, belongs to the technical field of adverse drug reaction prediction, and solves the problem of inaccurate prediction in the prior art. The method comprises the following steps: acquiring characteristic data and observation data of drugs and adverse reactions, and constructing a training sample set; a joint learning model is constructed, and the joint learning model comprises a tendency score model used for predicting the probability that the drug-adverse reaction is observed; the adverse reaction prediction model is used for predicting the probability of adverse reaction of the drug; the delay feedback prediction model is used for predicting delay feedback time of the adverse reaction; training the joint learning model based on the constructed training sample set to obtain a trained joint learning model; and inputting the characteristic data of the to-be-predicted drug and the adverse reaction into the trained joint learning network to predict whether the drug causes the adverse reaction or not. And accurate unbiased drug adverse reaction prediction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of adverse drug reaction prediction, and in particular to a method and system for predicting adverse drug reactions with delayed feedback. Background Art

[0002] Adverse drug reactions (also called side effects) refer to the harmful effects of drugs on patients during routine treatment. To date, thousands of types of adverse reactions have been reported, many of which have led to serious unexpected and harmful consequences. Timely and effective ADR warnings can help standardize and guide the production of drugs with limited side effects. In the problem of predicting adverse drug reactions, the same drug may cause multiple adverse reactions, but the number of types of adverse reactions is much larger than the number of adverse drug reactions that a single drug may cause, resulting in sparsity in the data of adverse drug reactions. Many previous methods are based on association rule mining and statistical significance tests to identify significant associations between drugs and ADRs, but their performance in prediction tasks is limited. Another class of deep learning-based methods uses pharmacovigilance networks, ensemble methods, and deep neural network models to achieve good predictive performance in predicting specific adverse reactions. However, since drugs may cause multiple ADRs at the same time, directly reusing single-label prediction methods without considering the deep connections between drugs-drugs, ADR-ADRs, and drug-ADRs may result in time-consuming and poor prediction performance.

[0003] Accurately predicting the probability of a drug producing a specific adverse reaction can guide the production of new drugs and minimize the probability of (serious) adverse reactions.

[0004] In practical applications, there may be a lack of high-quality explicit feedback. On the contrary, there is a large amount of implicit feedback in the data. For example, in the drug adverse reaction system, the adverse reactions caused by the drug are usually collected by users, which cannot capture all the adverse reactions of each drug. One of the most serious problems with implicit feedback is the existence of selection bias. In other words, the collected drug-adverse reaction pairs are usually a set of "non-random missing" samples of the overall sample: in the adverse reaction reporting system, more serious adverse reactions have a greater probability of being reported, while milder adverse reactions are easily ignored. Therefore, directly exploring the relationship between drugs and adverse reactions from the observed drug-adverse reaction pairs may lead to a decrease in the performance of the drug-adverse reaction prediction model. For example, for drugs for the treatment of cancer or leukemia, some serious adverse reactions will be reported (such as shock or heavy bleeding), while some milder adverse reactions (such as stomach pain, dizziness) will not be reported even if they occur, resulting in these adverse reactions that occurred but not reported and adverse reactions that did not occur being negative samples in the observed data, thus leading to the problem of selection bias.

[0005] In order to solve the selection bias of implicit feedback, traditional feature engineering-based methods collect feature information of drugs and adverse reactions, and use the combination of these features to predict adverse reactions of drugs. In addition, some collaborative filtering methods based on matrix factorization (MF) and neural matrix factorization (NMF) can capture the potential connection between drugs and adverse reactions. These methods have been widely used to deal with implicit feedback. However, previous methods ignored the problem of delayed feedback, which often occurs in real scenarios. In the prediction of adverse drug reactions, it takes a certain amount of time for an individual to have an adverse reaction after taking a certain drug, and the conversion time of this part may be too long and exceed the observation time of the adverse reaction reporting system, resulting in these drug adverse reaction pairs that should be positive samples being marked as negative samples in the system, resulting in false negative samples, further causing sample bias. Ignoring delayed feedback when predicting adverse drug reactions will lead to inaccurate predictions. Summary of the invention

[0006] In view of the above analysis, the embodiments of the present invention aim to provide a method and system for predicting adverse drug reactions with delayed feedback, so as to solve the problem of inaccurate prediction of adverse drug reactions when there is both selection bias and delayed feedback.

[0007] In one aspect, an embodiment of the present invention provides a method for predicting adverse drug reactions with delayed feedback, comprising the following steps: Obtain characteristic data and observation data of drugs and adverse reactions, and construct a training sample set; Constructing a joint learning model, the joint training model includes: a propensity score model for predicting the probability of drug-adverse reactions being observed; an adverse reaction prediction model for predicting the probability of adverse reactions caused by drugs; and a delayed feedback prediction model for predicting the delayed feedback time of adverse reactions; Training the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The characteristic data of the drug to be predicted and the adverse reaction are input into the trained joint learning network to predict whether the drug will cause the adverse reaction.

[0008] Based on the further improvement of the above method, the following formula is used to calculate the loss of the joint learning model: ; in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix.

[0009] Based on the further improvement of the above method, the following formula is used to calculate the prediction loss: ; in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

[0010] Based on the further improvement of the above method, the propensity score model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, represents the parameters of the propensity score model.

[0011] Based on the further improvement of the above method, the adverse reaction prediction model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the adverse reaction prediction model.

[0012] Based on the further improvement of the above method, the delayed feedback prediction model uses the following formula to model the delayed feedback time of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the delayed feedback prediction model.

[0013] On the other hand, an embodiment of the present invention provides a drug adverse reaction prediction system with delayed feedback, comprising: The sample set construction module is used to obtain the characteristic data and observation data of drugs and adverse reactions and construct the training sample set; A model building module is used to build a joint learning model, wherein the joint training model includes: a propensity score model, which is used to predict the probability of drug-adverse reactions being observed; an adverse reaction prediction model, which is used to predict the probability of a drug producing an adverse reaction; and a delayed feedback prediction model, which is used to predict the delayed feedback time of an adverse reaction; A model training module, used to train the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The prediction module is used to input the characteristic data of the drug to be predicted and the adverse reaction into the trained joint learning network to predict whether the drug will cause the adverse reaction.

[0014] Based on the further improvement of the above system, the following formula is used to calculate the loss of the joint learning model: ; in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix.

[0015] Based on the further improvement of the above system, the following formula is used to calculate the prediction loss: ; in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

[0016] Based on the further improvement of the above system, the adverse reaction prediction model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the adverse reaction prediction model.

[0017] Compared with the prior art, the present invention constructs a training sample set by acquiring characteristic data and observation data of drugs and adverse reactions, and constructs a joint learning model including a propensity score model, an adverse reaction prediction model and a delayed feedback prediction model. The influence of selective bias is mitigated by the propensity score model, so that the final learned prediction model is more robust in the face of selective bias. The accuracy of adverse reaction prediction is improved by multi-task learning that simultaneously optimizes adverse reaction prediction and delayed feedback prediction, and unbiased adverse reaction prediction is provided in scenarios with both selection bias and delayed feedback, thereby further improving the prediction accuracy.

[0018] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components; Figure 1 A flowchart of a method for predicting adverse drug reactions with delayed feedback according to an embodiment of the present invention; Figure 2 The block diagram of the drug adverse reaction prediction system with delayed feedback according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0021] A specific embodiment of the present invention discloses a method for predicting adverse drug reactions with delayed feedback, such as Figure 1 As shown, the following steps are included: Obtain characteristic data and observation data of drugs and adverse reactions, and construct a training sample set; Constructing a joint learning model, the joint training model includes: a propensity score model for predicting the probability of drug-adverse reactions being observed; an adverse reaction prediction model for predicting the probability of adverse reactions caused by drugs; and a delayed feedback prediction model for predicting the delayed feedback time of adverse reactions; Training the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The characteristic data of the drug to be predicted and the adverse reaction are input into the trained joint learning network to predict whether the drug will cause the adverse reaction.

[0022] Compared with the prior art, the method for predicting adverse drug reactions with delayed feedback provided in this embodiment constructs a training sample set by acquiring characteristic data and observation data of drugs and adverse reactions, and constructs a joint learning model including a propensity score model, an adverse reaction prediction model and a delayed feedback prediction model. The influence of selective bias is mitigated by the propensity score model, so that the final learned prediction model is more robust in the face of selective bias. The accuracy of adverse reaction prediction is improved by multi-task learning that simultaneously optimizes adverse reaction prediction and delayed feedback prediction, and unbiased adverse reaction prediction is provided in scenarios with both selection bias and delayed feedback, thereby further improving prediction accuracy.

[0023] The method of the present invention can accurately predict adverse drug reactions and help identify possible side effects early, thereby reducing the risk of serious adverse reactions after the drug is marketed. This helps protect the health of patients and reduce medical accidents caused by drugs. By predicting adverse reactions in the early stages of drug development, pharmaceutical companies can more effectively screen and optimize candidate drugs and reduce the risk of failure. This can save research and development costs and time and improve the efficiency of drug development. Accurately predicting adverse drug reactions before clinical trials can help design safer test plans, select appropriate dosage ranges and subjects, thereby reducing risks in trials and improving the success rate of trials. In addition, by predicting adverse drug reactions in different individuals, it can help doctors formulate personalized medication plans based on individual characteristics of patients (such as genetic information, previous medical history, etc.), thereby improving treatment effects and reducing adverse reactions.

[0024] In the following, capital letters are used to represent random variables, and lowercase letters are used to represent the corresponding data. The variables are defined as follows (for simplicity, the subscript i is omitted in some formulas): X represents a feature set, including the characteristics of drugs and adverse reactions. Represents the characteristic data of the i-th drug-adverse reaction pair.

[0025] Indicates whether to report this drug-adverse reaction pair.

[0026] Represents the true label of a drug-adverse reaction pair.

[0027] : Indicates the time from taking the medicine to the occurrence of adverse reactions, that is, the delayed feedback time.

[0028] :Observation time after taking the medicine.

[0029] Indicates whether adverse reactions occurred during the observation time.

[0030] A drug and an adverse reaction constitute a drug-adverse reaction pair.

[0031] Observational data include whether the ith drug-adverse reaction pair is reported , the true mark , observation time , Observation outcome , Delayed time of adverse reactions .

[0032] During implementation, the characteristic data of the drug-adverse reaction pair can be extracted based on the drug information and the adverse reaction information. For example, the drug and the adverse reaction are coded respectively, and the drug code and the adverse reaction code are combined to form the characteristic data of the drug-adverse reaction pair.

[0033] Unlike the prior art, the present invention does not require the assumption that In the case of and are independent of each other. At the same time, the present invention assumes and the observation time G. This assumption is valid because the observation time is determined by the characteristics of the patient (such as the patient's stay in the hospital), while whether an adverse reaction occurs and when it occurs are determined by the essential characteristics of the drug and the patient.

[0034] From the dataset, we can obtain the time for the observed drug-adverse reaction pairs to produce adverse reactions (the time to complete the conversion) , characteristic data and observation time . Let the entire data set be , the set of observed drug-adverse reaction pairs is The observed data consists of features , observation outcome , observation time For the observed outcomes In the case of adverse reactions, the delayed feedback time from taking the medicine to the occurrence of adverse reactions can be further obtained. .

[0035] Therefore, the characteristic data and observation data of drugs and adverse reactions are obtained, and each sample in the constructed training sample set includes the characteristic data of the drug-adverse reaction pair , observation time , Observation outcome , observation outcome The positive samples also include the delayed time of adverse reactions .

[0036] In implementation, the present invention assumes that the probability distribution of the delay time follows an exponential distribution, and the risk function of delayed feedback is expressed as Under the exponential distribution, due to the delayed feedback factor, for the characteristic The probability that an adverse drug reaction is not observed is .

[0037] During implementation, in order to improve the accuracy of adverse drug reaction prediction, the present invention constructs a joint learning model, which includes a propensity score model, an adverse reaction prediction model and a delayed feedback prediction model.

[0038] During implementation, in order to solve the problem of selection bias, the joint learning model of the present invention includes a propensity score model for predicting the probability of drug-adverse reactions being reported.

[0039] During implementation, a logistic regression model was used to construct the propensity score model.

[0040] Specifically, the propensity score model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, represents the parameters of the propensity score model. That is, the probability of the ith drug-adverse reaction being reported calculated by the propensity score model.

[0041] During implementation, the probability of the i-th drug-adverse reaction being reported is calculated using formula (1).

[0042] During implementation, in order to further fit the observed data, the present invention uses exponential distribution modeling to delay the feedback time.

[0043] During implementation, the delayed feedback prediction model is constructed using exponential distribution; the delayed feedback prediction model uses the following formula to calculate the delayed feedback time of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Parameters representing delayed feedback prediction. That is, the exponential distribution parameter corresponding to the delayed feedback time of the i-th drug-adverse reaction pair predicted by the delayed feedback prediction model.

[0044] Adverse reaction prediction model is used to predict the probability of adverse reactions of drugs.

[0045] During implementation, a logistic regression model is used to construct the adverse reaction prediction model; the adverse reaction prediction model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters for adverse reaction prediction. That is, the probability of adverse reaction of the i-th drug-adverse reaction pair calculated by the adverse reaction prediction model.

[0046] The present invention constructs a joint learning model and performs multi-task learning so that multiple tasks assist each other. When there is delayed feedback and selective bias, the accuracy of adverse drug reaction prediction can be improved.

[0047] During implementation, the present invention uses the following formula to calculate the loss of the joint learning model: ; in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix. is the negative log-likelihood function.

[0048] In order to prevent overfitting and improve the stability of the model, the training loss of the joint learning model Regularization terms for the propensity score model, adverse reaction prediction model, and delayed feedback prediction model are added .

[0049] For the ith drug-adverse reaction pair, if the drug is finally observed to produce the adverse reaction, that is, , which means that the drug-adverse reaction pair was reported, i.e. , and the true marker for the drug-adverse reaction pair is positive The time to observe adverse reactions is , , represents the observation time of the ith drug-adverse reaction pair.

[0050] For the ith drug-adverse reaction pair, if the adverse reaction is not observed in the end, that is, , there are three situations: the drug does not have the adverse reaction, that is ; The drug has this adverse reaction, but it is not reported ( , ); The drug has the adverse reaction and is reported, but no adverse reaction occurs within the observation time, resulting in a negative sample being reported ( , , ).

[0051] Therefore, for The probability of observing adverse drug reactions during the observation time is the product of the probability of observing a drug-adverse reaction pair, the probability of the drug producing an adverse reaction, and the probability that the time of the adverse reaction is less than or equal to the observation time. Taking the logarithm, we get the loss of the positive sample. .

[0052] for For a sample, the probability that no adverse drug reaction is observed during the observation time is the sum of the following three probabilities: no drug-adverse reaction pair is observed; drug-adverse reaction pair is observed but no adverse reaction occurs; drug-adverse reaction pair is observed and adverse reaction occurs, but the time of adverse reaction exceeds the observation time. Therefore, for negative samples, the loss is .

[0053] Therefore, the prediction loss is calculated using the following formula: ; in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

[0054] When the traditional likelihood function is directly maximized on the sample, the prediction result may be biased due to the existence of selective bias and delayed feedback. The present invention improves the likelihood function and optimizes the improved likelihood function through multi-task learning to simultaneously handle selective bias and delayed feedback, thereby obtaining a more accurate adverse drug reaction prediction model.

[0055] A specific embodiment of the present invention discloses a drug adverse reaction prediction system with delayed feedback, such as Figure 2 As shown, the system includes: The sample set construction module is used to obtain the characteristic data and observation data of drugs and adverse reactions and construct the training sample set; A model building module is used to build a joint learning model, wherein the joint training model includes: a propensity score model, which is used to predict the probability of drug-adverse reactions being observed; an adverse reaction prediction model, which is used to predict the probability of a drug producing an adverse reaction; and a delayed feedback prediction model, which is used to predict the delayed feedback time of an adverse reaction; A model training module, used to train the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The prediction module is used to input the characteristic data of the drug to be predicted and the adverse reaction into the trained joint learning network to predict whether the drug will cause the adverse reaction.

[0056] Preferably, the loss of the joint learning model is calculated using the following formula: ; in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix.

[0057] Preferably, the prediction loss is calculated using the following formula: ; in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

[0058] Preferably, the adverse reaction prediction model uses the following formula to calculate the prediction result of the i-th drug-adverse reaction pair: ; in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the adverse reaction prediction model.

[0059] The above method embodiments and system embodiments are based on the same principle, and their related parts can be used for reference, and can achieve the same technical effect. The specific implementation process refers to the above embodiment, which will not be repeated here.

[0060] Electronic device embodiment: A specific implementation of the present application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method for predicting adverse drug reactions with delayed feedback in the method embodiment when executing the program.

[0061] Readable storage medium embodiment: A specific implementation of the present application discloses a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method for predicting adverse drug reactions with delayed feedback in the method embodiment are implemented.

[0062] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0063] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting adverse drug reactions with delayed feedback, characterized in that: The following steps are involved: Obtain characteristic data and observation data of drugs and adverse reactions, and construct a training sample set; Constructing a joint learning model, the joint learning model includes: a propensity score model for predicting the probability of drug-adverse reactions being observed; an adverse reaction prediction model for predicting the probability of drug-adverse reactions; and a delayed feedback prediction model for predicting the delayed feedback time of adverse reactions; Training the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The characteristic data of the drug to be predicted and the adverse reaction are input into the trained joint learning network to predict whether the drug will cause the adverse reaction.

2. The method for predicting adverse drug reactions with delayed feedback according to claim 1, characterized in that: The loss of the joint learning model is calculated using the following formula: ； in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix.

3. The method for predicting adverse drug reactions with delayed feedback according to claim 2, characterized in that: The prediction loss is calculated using the following formula: ； in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

4. The method for predicting adverse drug reactions with delayed feedback according to claim 3, characterized in that: The propensity score model uses the following formula to calculate the prediction result of the ith drug-adverse reaction pair: ； in, represents the characteristic data of the ith drug-adverse reaction pair, represents the parameters of the propensity score model.

5. The method for predicting adverse drug reactions with delayed feedback according to claim 3, characterized in that: The adverse reaction prediction model uses the following formula to calculate the prediction result of the ith drug-adverse reaction pair: ； in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the adverse reaction prediction model.

6. The method for predicting adverse drug reactions with delayed feedback according to claim 3, characterized in that: The delayed feedback prediction model uses the following formula to model the delayed feedback time of the ith drug-adverse reaction pair: ； in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the delayed feedback prediction model.

7. A drug adverse reaction prediction system with delayed feedback, characterized in that: include: The sample set construction module is used to obtain the characteristic data and observation data of drugs and adverse reactions and construct the training sample set; A model building module is used to build a joint learning model, wherein the joint learning model includes: a propensity score model, which is used to predict the probability of drug-adverse reactions being observed; an adverse reaction prediction model, which is used to predict the probability of adverse reactions caused by drugs; and a delayed feedback prediction model, which is used to predict the delayed feedback time of adverse reactions; A model training module, used to train the joint learning model based on the constructed training sample set to obtain a trained joint learning model; The prediction module is used to input the characteristic data of the drug to be predicted and the adverse reaction into the trained joint learning network to predict whether the drug will cause the adverse reaction.

8. The drug adverse reaction prediction system with delayed feedback according to claim 7, characterized in that: The loss of the joint learning model is calculated using the following formula: ； in, represents the prediction loss, represents the parameters of the propensity score model, represents the parameters of the adverse reaction prediction model, represents the parameters of the delayed feedback prediction model, represents the regularization parameter, Represents the 2-norm of a matrix.

9. The drug adverse reaction prediction system with delayed feedback according to claim 8, characterized in that: The prediction loss is calculated using the following formula: ； in, represents the prediction result of the propensity score model for the ith drug-adverse reaction pair, represents the prediction result of the adverse reaction prediction model for the i-th drug-adverse reaction pair, represents the prediction result of the delayed feedback prediction model for the i-th drug-adverse reaction pair, represents the delayed feedback time of the ith drug-adverse reaction pair; represents the observation time of the i-th drug-adverse reaction pair; represents the characteristic data of the ith drug-adverse reaction pair, Indicates that the i-th drug-adverse reaction pair is a positive sample, Indicates that the i-th drug-adverse reaction pair is a negative sample.

10. The drug adverse reaction prediction system with delayed feedback according to claim 9, characterized in that: The adverse reaction prediction model uses the following formula to calculate the prediction result of the ith drug-adverse reaction pair: ； in, represents the characteristic data of the ith drug-adverse reaction pair, Represents the parameters of the adverse reaction prediction model.

Images