Machine Learning Based Adverse Drug Reaction Extraction From Text

Monitor the adverse drug reactions from text is not only an important way to ensure patients’ safety of using drug,but also a significant mean of pharmacovigilance.Traditional methods,such as spontaneous reporting system has disadvantage of inefficiencys.As for natural language processing,it takes much time and needs specific knowledge to construct rules.To solve these problems,more and more attentions are focused on utilizing machine learning to extract the adverse drug reaction from text.How to improve the performance of machine learning is a meaningful subject.Specifically,the main contents include:(1)Feature selection.Feature sets are constructed from corpus of text based on the lexicon-matched,N-gram and topic model.Chi-square test and wrapper method are chosen to select best feature sets.(2)Adverse drug reaction extraction from text based on single classifier.Five kinds of machine learning models commonly used in text classification task are adopted:Naive Bayes,Logistic Regression,Support Vector Machine,Decision Tree and Random Forest.Through the ten-fold cross-validation,the parameters of these classifiers are tuned.Besides,the generalization performance of these classifiers are compared according to their F1.(3)Adverse drug reaction extraction from text based on ensemble classifiers.Based on the Bagging,plurality voting,weighted voting,threshold-moving and weighted averaging strategies are separately employed to integrate these above classifiers.The performance of ensemble classifiers are compared through multiple experiments.The experimental results show that by using the feature set constructed in this paper,Logistic Regression classifier’s F1(80.33%)is higher 3.33%than the previous best research from the same text.Moreover,the F1(82.59%)of the weighted averaging ensemble classifiers increases by 5.59%.