Adverse Drug Reaction Monitoring Based On Text Classification

Adverse drug reaction monitoring is one of the important means to insure drug safety.How to effectively monitor adverse drug reaction is crucial to public health and has significant research value.The main methods for monitoring adverse drug reaction are clinical trials and spontaneous adverse drug reaction report,both of which are costly and inefficient.In recent years,the emerged text classification based monitoring method has gradually become a research hotspot.It can make full use of the existing data related to adverse drug reaction and automatically monitor the adverse reaction after drug flowing to the market.However,the following drawbacks exist:1)The data sources of adverse drug reaction monitoring based on text classification are mainly medical case reports,etc.The text length is short and the features are sparse.As a result,the combination of medical knowledge and statistical features is needed,and the classification performance still needs to be improved;2)The method based on convolutional neural network model was initially applied to adverse drug reaction monitoring,but most of them use word embedding pre-trained on the general domain corpus,to represent adverse drug reaction texts,and less use word embedding pre-trained on medical literature corpus.The main work of this paper is as follows:?1?Monitoring of adverse drug reaction based on combined features.Three different features are used to combine medical domain knowledge with the characteristics of adverse drug reaction texts,including medical dictionary,N-Gram,topic model.Then three types of features are combined,the combinations including medical dictionary+N-Gram,medical dictionary+topic model,N-Gram+topic model,medical dictionary+N-Gram+topic model.Finally,the classification performance of different combination features on naive Bayes,logistic regression,support vector machine and random forest are compared and analyzed.?2?Monitoring of adverse drug reaction based on convolutional neural networks.The convolutional neural network model can automatically extract features and classify text.This paper constructs a convolutional neural network model,extracts local features through multi-convolution kernel,and integrates medical domain knowledge with word embedding pre-trained on medical literature corpus.Finally,the classification performance of the adverse drug reaction task is compared and analyzed.The effectiveness of the proposed method is validated on the medical case report dataset ADE.The experimental results show that the classification performance based on the combination of medical dictionary+N-Gram+topic model is optimal on the support vector machine model,and the value of₁ is 83.54%.Compared with the current known best result,the value of₁ is increased by 2.11%.The classification performance₁ value of the convolutional neural network model based on medical corpus word vector training is 88.31%,and the value of₁ is improved by 4.77%compared with the optimal result based on the combined feature;and the value of₁ was increased by 1.31%compared with the current known best result based on the convolutional neural network.