Detecting Adverse Drug Reactions Using a Sentiment Classification Framework

Medical blogs and forums are a source of sentiment oriented content that is used in diverse applications including post-marketing drug surveillance, competitive intelligence and the assessment of health-related opinions and sentiments for detecting adverse drug reactions. However applying existing tools for sentiment analysis to health-related datasets provides inadequate classification accuracy. These methods employ less useful features sets and therefore lack discriminatory potential. In this study we propose a framework that uses feature set ensembles with novel feature representations that reduce sparsity by adding representational richness. Our framework extracts important semantic, sentiment, and affect cues, that are better able to reflect the experiences of people when they discuss adverse drug reactions as well as the severity and the emotional impact of their experiences. Experiments conducted on a test bed of health-2.0 datasets, demonstrate improved classification accuracy in comparison to existing techniques. Furthermore, the proposed framework is able to detect adverse drug events earlier, and with higher recall than comparison methods, thereby demonstrating its utility for social media based post-marketing drug surveillance.