An attention-based effective neural model for drug-drug interactions extraction

Area of Science:

• Pharmacovigilance
• Natural Language Processing
• Computational Linguistics

Background:

• Drug-drug interactions (DDIs) pose risks to patient safety and increase healthcare costs.
• Current text-mining systems struggle to accurately classify DDIs within long and complex sentences.
• Improved DDI classification is essential for clinical recognition and management.

Purpose of the Study:

• To develop an effective model for classifying drug-drug interactions (DDIs) from biomedical literature.
• To enhance the accuracy of DDI classification, particularly in challenging long and complex sentences.
• To improve the clinical recognition of DDIs for better patient safety and cost control.

Main Methods:

• A novel model combining an attention mechanism with a long short-term memory (LSTM) recurrent neural network.
• Candidate-drug-oriented input attention to identify influential words for drug pairs.
• Integration of position- and POS-embedding vectors with bidirectional LSTM for semantic information extraction.
• Softmax layer for final DDI classification.

Main Results:

• The proposed system achieved superior performance in DDI detection (84.0%) and classification (77.3%) on the DDIExtraction 2013 corpus.
• On the Medline-2013 dataset, the system's F-score significantly surpassed top-ranking systems by 12.6% for long and complex sentences.
• The model demonstrated improved recognition of both close-range and long-range word patterns.

Conclusions:

• The developed model significantly enhances DDI classification performance.
• The approach effectively addresses the challenge of classifying DDIs in long, complex, and compound sentences.
• This advancement contributes to more accurate identification of drug-drug interactions in clinical settings.