Classifying adverse drug reactions from imbalanced twitter data.

Area of Science:

• Pharmacovigilance and Drug Safety
• Computational Linguistics
• Machine Learning for Healthcare

Background:

• Social media is increasingly used for public health information sharing, including adverse drug reactions (ADRs).
• Mining social media for ADRs aids post-marketing surveillance but faces challenges from highly imbalanced datasets, hindering classifier performance.
• Existing methods for imbalanced data lack investigation in the context of social media ADR detection.

Purpose of the Study:

• To evaluate various imbalanced classification techniques for ADR detection from social media data.
• To propose and assess a novel word embedding-based synthetic minority over-sampling technique (WESMOTE).

Main Methods:

• Comparison of multiple imbalanced classification strategies on large, imbalanced datasets for ADR detection.
• Development and evaluation of WESMOTE, synthesizing examples from word embedding-based sentence representations.
• Performance assessment using F-scores against state-of-the-art approaches.

Main Results:

• Classifiers using imbalanced techniques achieved comparable or superior F-scores compared to state-of-the-art methods.
• Optimal configurations integrated random under-sampling with WESMOTE, boosting, or ensemble methods.
• Ensemble methods like Vote-based Under-sampling (VUE) and random under-sampling boosting showed promise as alternatives to hybrid synthetic methods.

Conclusions:

• Word embedding-based synthetic over-sampling (WESMOTE) is more effective than traditional methods for imbalanced text classification.
• Combining imbalanced strategies with under-sampling techniques is preferred for large social media datasets.
• Simpler methods like VUE can achieve high performance with advantages in speed and ease of development.