Predicting Adverse Drug Reactions from Social Media Posts: Data Balance, Feature Selection and Deep Learning

Area of Science:

• Pharmacovigilance
• Computational Linguistics
• Machine Learning

Background:

• Social media provides valuable patient insights for post-marketing surveillance of adverse drug reactions (ADRs).
• Deriving predictive models from social posts faces challenges like data sparseness, high dimensionality, and term diversity.
• Existing methods require significant manual effort for feature engineering.

Purpose of the Study:

• To develop and evaluate an automated approach for identifying ADRs from social media data.
• To address data challenges inherent in social media text for ADR prediction.
• To enhance the predictive performance of ADR models using deep learning.

Main Methods:

• Data analytics focusing on data balance, feature selection, and feature learning.
• Comprehensive experimental analysis of various data processing and modeling techniques.
• Implementation of a deep learning model utilizing BERT (Bidirectional Encoder Representations from Transformers) with a batch-wise adaptive strategy.

Main Results:

• Both manual and automated feature engineering methods proved effective for ADR prediction.
• The proposed deep learning approach, leveraging BERT, demonstrated enhanced predictive performance.
• Automated feature learning significantly reduced manual effort compared to traditional machine learning methods.

Conclusions:

• Deep learning, particularly BERT with adaptive strategies, offers a powerful solution for ADR prediction from social media.
• Automated feature learning streamlines the process, making ADR surveillance more efficient.
• The study highlights the potential of social media data and advanced NLP techniques in pharmacovigilance.