Portable automatic text classification for adverse drug reaction detection via multi-corpus training

Area of Science:

• Pharmacovigilance and computational linguistics.
• Natural Language Processing (NLP) applications in healthcare.

Background:

• Automatic detection of adverse drug reactions (ADRs) from text is crucial for pharmacovigilance.
• Social media provides a vast, yet challenging, source of user-generated health data.
• Existing methods require optimization for accurate ADR identification.

Purpose of the Study:

• To explore NLP techniques for feature generation and machine learning for ADR classification.
• To introduce two novel datasets for ADR detection from social media.
• To investigate the impact of combining training data from distinct corpora on classification accuracy.

Main Methods:

• Utilized a feature-rich text classification approach incorporating semantic properties like sentiment and topic.
• Developed and annotated two in-house datasets from social media posts, alongside a clinical report dataset.
• Combined training data from different corpora to enhance classification performance.

Main Results:

• Achieved superior ADR detection F-scores (0.812) compared to previous benchmarks (0.770).
• Demonstrated significant improvements in F-scores for in-house datasets when combining corpora (up to 0.704).
• Showcased the effectiveness of semantic features and multi-corpus training.

Conclusions:

• Advanced NLP techniques and rich feature generation substantially improve ADR classification accuracy.
• Incorporating semantic features (topics, sentiments, polarities) enhances model performance.
• Multi-corpus training is beneficial, especially for imbalanced datasets like social media data, potentially reducing annotation costs.