Detecting Adverse Drug Events in Social Media: A Brief Literature Review

Area of Science:

• Pharmacovigilance and Public Health
• Computational Linguistics and Artificial Intelligence
• Digital Health and Social Media Analytics

Background:

• Adverse drug events (ADEs) pose significant public health challenges, necessitating innovative surveillance methods.
• Social media offers a rich, real-time source of patient-generated data for pharmacovigilance.
• Traditional pharmacovigilance methods can be complemented by analyzing online discourse.

Purpose of the Study:

• To systematically review recent natural language processing (NLP) research focused on identifying ADEs within social media text.
• To synthesize findings from 100 peer-reviewed studies published between 2017 and 2025.
• To identify trends, recurrent challenges, and future directions in NLP for social media-based ADE surveillance.

Main Methods:

• Systematic literature review of 100 peer-reviewed studies identified via Google Scholar.
• Exclusion of patents, protocols, and conference abstracts lacking methodological detail.
• Categorization of studies into classification, extraction, normalization, corpus creation, and analytical tasks, with data extraction on objectives, data sources, preprocessing, models, and evaluation metrics.

Main Results:

• Twitter is the predominant data source (78%), with English being the dominant language (approx. 95%).
• Transformer-based models, particularly BERT variants (e.g., BERTweet, BioBERT), are used in over 60% of studies.
• Key challenges include noisy/imbalanced data, multilingual/code-mixed content, implicit drug-event expressions, and a lack of comprehensive annotation guidelines (12% of papers).

Conclusions:

• NLP techniques show promise for ADE surveillance in social media, but challenges in data quality, multilingual support, and model deployment persist.
• Transformer models have become standard, yet issues like class imbalance and implicit ADE mentions require further attention.
• Future research should focus on model explainability, low-resource languages, multimodal data, and closer collaboration with regulatory bodies for real-world application.