Adverse Drug Reactions Detection from Social Media: an Empirical Evaluation of Machine Learning Techniques

Abstract

This research addresses the critical concern of Adverse Drug Reactions (ADRs), emphasizing the need for their timely detection to safeguard patient well-being. Detecting ADRs within sentences is pivotal for effective public health monitoring. The study's primary objective is to assess whether sentences contain ADR references, a crucial step in identifying potential ADRs early. Timely recognition can mitigate patient harm and enhance drug development processes. The rise of patient engagement on social media has turned it into a valuable real-time resource for ADR-related information. Patients increasingly share their personal narratives about drug usage on these platforms, creating an innovative avenue for gathering firsthand accounts of ADRs. This evolving trend has transformed social media into an indispensable source of information. The study aims to compare machine learning algorithms in classifying sentences as containing ADRs or not. Three diverse datasets - CADEC, TwiMed (PubMed), and ADE - are used to train and evaluate models. Rigorous experimentation highlights the superiority of the Naive Bayes classifier over other methods. Notably, this classifier achieves remarkable accuracy rates of 94.29%, 78.76%, and 64.93%, on the CADEC, ADE, and PubMed datasets, respectively. This comparative study demonstrates the effectiveness of machine learning in identifying ADRs within sentences and underscores the Naive Bayes classifier's consistently impressive performance across different datasets.