Adverse Drug Reaction Detection Using Latent Semantic Analysis

Detecting Adverse Drug Reactions (ADRs) is one of the important information for determining the view of the patient on one drug. Most studies have investigated the extraction of ADRs from social networks, in which users share their opinion on a particular medication. Some studies have used trigger terms to detect ADRs. Such studies showed remarkable performance in terms of extracting ADR. However, these terms only would not be sufficient since it needs to be extended periodically when new side effects or new medical-related entities are being discovered. In addition, the feature space with trigger terms would lack latent semantic. This study aims to propose a semantic method based on Latent Semantic Analysis (LSA) for improving the detection of ADR. A benchmark dataset has been used in the experiments along with several pre-processing operations that have been applied including stop word removal, tokenization and stemming with three classifiers that were trained on the proposed LSA, namely Support Vector Machine (SVM), Naïve Bayes (NB) and Linear Regression (LR). In addition, two representations of documents were used, namely Term Frequency (TF) and Term Frequency-Inverse Document Frequency (TF-IDF). Results showed that the proposed LSA outperformed the baseline extended trigger terms by achieving 82% of F-measure for the dataset. Such superiority highlights the use of LSA where the semantic correspondences could be identified correctly rather than using a predefined list of trigger terms.