Hierarchical feature modeling with data augmentation and focal loss for drug–drug interaction extraction

Abstract

Drug–drug interaction (DDI) extraction is crucial for ensuring medication safety but faces challenges from complex biomedical text structures and data imbalance. To address these issues, we propose HiFAB-DDI (Hierarchical Features with Augmentation and Focal-Attention-Based Loss), a novel architecture that combines CNNs for capturing local dependencies and Transformers for modeling global contexts. Additionally, HiFAB-DDI leverages BioGPT-based data augmentation to enrich training data with diverse and domain-specific examples and incorporates an attention-enhanced focal loss mechanism to handle data imbalance effectively. Our method demonstrates state-of-the-art performance on benchmark datasets, achieving significant improvements in precision, recall, and F1 Score compared to existing approaches. Specifically, HiFAB-DDI achieves an F1 Score of 84.78%, outperforming baseline models by a significant margin. Ablation studies confirm the critical contributions of data augmentation, hierarchical modeling, and focal loss with attention. Our code and data are publicly available at: https://github.com/Hero-Legend/HiFAB-DDI.