A Gated Multi-hop Attention Fusion Network for Extracting Inter-Sentential Adverse Drug Event Relation

Area of Science:

• Pharmacovigilance and Drug Safety
• Biomedical Natural Language Processing
• Machine Learning for Healthcare

Background:

• Automatic extraction of inter-sentential relations between adverse drug event (ADE) entities is crucial for pharmacovigilance.
• Current deep learning models face challenges in simultaneously leveraging entity, sentence, and document-level information for efficient representation learning.
• The sparsity of relevant entity pairs with predefined relations poses a significant problem for model training.

Purpose of the Study:

• To propose a novel Gated Multi-hop Attention Fusion Network (GMAFNet) for improved extraction of inter-sentential ADE relations.
• To address the limitations of existing methods in representation learning and the sparsity problem.
• To enhance the efficiency and accuracy of ADE relation extraction from unstructured documents.

Main Methods:

• Feature extraction at three levels: entity-level (adaptive localized context pooling), sentence-level (attentive hierarchical context pooling), and document-level (attentive global-local aggregation context pooling).
• Fusion of extracted features using a novel gated multi-hop attention mechanism.
• Application of an asymmetric focal loss function to mitigate the issue of data sparsity.

Main Results:

• GMAFNet achieved a superior F1-score of 97.27% on the n2c2 2018 challenge benchmark dataset.
• The model demonstrated strong generalizability, reaching F1-scores of 78.19% on the CDR dataset and 87.10% on the GDA dataset.
• GMAFNet outperformed state-of-the-art systems in extracting complex inter-sentential ADE relations.

Conclusions:

• The proposed GMAFNet system is highly effective for extracting inter-sentential ADE relations.
• The multi-level feature fusion and asymmetric focal loss contribute significantly to the model's performance.
• GMAFNet represents a significant advancement in automated pharmacovigilance and drug safety analysis.