Mining Potential Adverse Drug Reactions Based On Social Media Data Analysis

In the 21st century,with the development of Web 2.0 technology,social media(such as forums and blogs)have gradually become popular among the crowd.These social media data produce a large amount of user-generated content.On medical and health social media,a large number of drug experience and comments from Internet users have been collected.These data contains information about the latest adverse drug reactions.The mining of potential adverse drug reactions can discover adverse reactions early by analyzing the use experience of drugs.It is of great significance to patients,pharmaceutical companies,and even the national health system.This paper systematically studies the methods of mining potential adverse drug reactions from social media data,including the recognition of adverse drug reaction entities,the normalization of adverse drug reaction entities,and the analysis of potential adverse drug reactions.The above methods are evaluated on the CADEC and Psy TAR datasets.Aiming at the entity recognition of adverse reactions,this paper firstly proposes an adverse reaction entity recognition method based on sequence annotation and establishes a benchmark model.Considering the long expression of adverse reactions mentioned in social media data and the connection between drugs and adverse reactions,based on the benchmark model,this paper proposes the adverse reaction entity recognition method that incorporate entity boundary information and drug information.The experience shows that the improved method proposed in this paper can achieve the highest F1 value of 71.32% and73.73% on the CADEC and Psy TAR datasets,respectively,which is 7.33% and9.42% higher than the benchmark model.Aiming at the entity normalization of adverse reactions,this paper proposes three methods,namely,the entity normalization method based on multiclassification,the entity normalization method based on sorting,and the entity normalization method based on seq2 seq.The experience shows that the seq2 seqbased entity normalization method performs the worst,and lacks the mapping process of adverse reaction entities to standard concepts,which is not very practical.The entity normalization method based on multi-classification has the problem of poor model generalization due to less training data.The sort-based entity normalization method solves the above-mentioned drawbacks by combining candidate generation and candidate sorting.The F1 value on CADEC and Psy TAR datasets reaches 61.73% and 65.83%,respectively.For the analysis of potential adverse drug reactions,this paper adopts the best-performing model in the adverse reaction entity recognition and entity normalization phase.This paper selects "liptor" as the target drug,and recognizes all the adverse reactions mentioned in the dataset.After filtering and comparing with the official statistics of adverse reactions,as well as the data demonstration in related research reports,the method for mining potential adverse drug reactions proposed in this paper has certain practicality and feasibility.