Detecting Personal Medication Intake in Twitter via Domain Attention-Based RNN with Multi-Level Features

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Pharmacovigilance

Background:

• Personal medication intake detection from social media is vital for drug safety surveillance.
• Existing models often neglect crucial medical domain information, limiting their effectiveness.
• Twitter data presents unique challenges for accurate medication consumption identification.

Purpose of the Study:

• To develop an advanced model for detecting personal medication intake from Twitter data.
• To integrate medical domain knowledge into a deep learning framework for improved accuracy.
• To enhance drug safety surveillance through more precise identification of medication consumption patterns.

Main Methods:

• Proposed a novel domain attention mechanism integrated with recurrent neural networks (LSTMs).
• Utilized character-level Convolutional Neural Networks (CNN) for morphological feature extraction.
• Employed a Bidirectional Long Short-Term Memory (BiLSTM) network with an attention mechanism over hidden states.
• Implemented multi-level feature representation of Twitter data for comprehensive analysis.

Main Results:

• The proposed model achieved a precision of 0.708, recall of 0.694, and F1 score of 0.697 on a public benchmark dataset.
• Demonstrated significant performance improvement over strong and relevant baseline models.
• Showcased the effectiveness of the domain attention mechanism in leveraging medical information.

Conclusions:

• The domain attention mechanism effectively incorporates medical domain information, leading to superior performance in personal medication intake detection.
• The multi-level feature representation approach enhances the model's ability to understand nuanced Twitter data related to medication.
• This research contributes to advancing automated drug safety surveillance systems through improved natural language processing techniques.