Evaluating knowledge fusion models on detecting adverse drug events in text

Area of Science:

• Pharmacovigilance and Natural Language Processing (NLP)

Background:

• Detecting adverse drug events (ADEs) is crucial for ongoing drug safety monitoring by regulatory bodies and the pharmaceutical industry.
• Social media platforms offer a rich source of real-world patient-reported adverse event data.
• Traditional methods require enhancement to effectively process diverse text sources for ADE detection.

Purpose of the Study:

• To evaluate knowledge fusion approaches combined with transformer-based NLP models for extracting ADE mentions.
• To assess the performance of these methods across various datasets, including social media, patient websites, and drug labels.
• To introduce and test a multi-modal architecture integrating transformer models and graph attention networks (GAT).

Main Methods:

• Formulated ADE extraction as a Named Entity Recognition (NER) task.
• Applied fusion learning to enhance transformer models with contextual knowledge from ontologies and knowledge graphs.
• Developed a multi-modal architecture combining transformer models (e.g., ERNIE, BERT) with GAT.

Main Results:

• Knowledge fusion models consistently outperformed the baseline BERT model across multiple corpora (PsyTAR, ADE, TAC).
• A multi-modality model (ERNIE + knowledge) achieved an F1-score of 71.84% on the CADEC corpus.
• A GAT and BERT combination achieved high F1-scores, including 94.15% on the TAC corpus.

Conclusions:

• Contextual knowledge significantly enhances the performance of knowledge fusion models for ADE detection.
• The proposed transformer-based and multi-modal approaches show strong potential for improving pharmacovigilance.
• These methods offer a robust framework for mining real-world data to ensure drug safety.