Deep neural networks ensemble for detecting medication mentions in tweets

Area of Science:

• Computational linguistics
• Pharmacoepidemiology
• Social media analytics

Background:

• Twitter is a valuable source of patient-generated data for public health insights.
• Accurate identification of medication mentions in tweets is crucial for pharmacoepidemiologic research.
• Traditional lexical searches for medication names are limited by misspellings and ambiguity.

Purpose of the Study:

• To develop an advanced method for automatically recognizing medication mentions in Twitter posts.
• To improve the recall and precision of identifying drug products and dietary supplements in tweets.

Main Methods:

• Introducing Kusuri, an Ensemble Learning classifier using four parallel modules: lexicon-based, spelling variant-based, pattern-based, and a weakly trained neural network.
• Employing a second module with an ensemble of deep neural networks to encode morphological, semantic, and long-range dependencies for final decision-making.
• Utilizing a two-stage approach to identify tweets potentially containing medication names and then making a final classification.

Main Results:

• Kusuri achieved an F1 score of 93.7% on a balanced corpus, nearing human annotator performance.
• On a highly imbalanced dataset (0.26% medication mentions), Kusuri obtained an F1 score of 78.8%, a novel achievement for such data.
• The system demonstrates high performance in identifying drug names within tweets.

Conclusions:

• The Kusuri system effectively identifies tweets mentioning drug names with high enough performance for practical application.
• This tool is ready for integration into public health surveillance systems, including pharmacovigilance and toxicovigilance pipelines.
• The developed method offers a significant advancement over traditional approaches for extracting medication information from social media.