Exploiting adversarial transfer learning for adverse drug reaction detection from texts

Area of Science:

• Pharmacovigilance and computational linguistics.
• Natural Language Processing (NLP) for healthcare.

Background:

• Adverse Drug Reactions (ADRs) pose significant risks to patient safety.
• Automated ADR detection is crucial for public health monitoring and pharmacovigilance.
• Existing ADR detection methods struggle with small-scale, diverse datasets and often require extensive manual annotation.

Purpose of the Study:

• To develop an effective ADR detection method for small target corpora.
• To address the challenges of noise and annotation workload in existing approaches.
• To leverage larger source corpora to improve performance on limited target data.

Main Methods:

• Formulating ADR detection as a text classification task.
• Introducing an adversarial transfer learning framework.
• Utilizing adversarial learning to minimize corpus-specific features and noise during knowledge transfer.

Main Results:

• The proposed adversarial transfer learning framework consistently outperformed state-of-the-art methods.
• Significant performance improvements were observed, particularly on small benchmark corpora.
• The method effectively mitigated noise introduced by combining corpora from different sources.

Conclusions:

• Adversarial transfer learning offers a robust solution for ADR detection in resource-limited settings.
• The framework enhances the utility of existing data for pharmacovigilance.
• This approach minimizes annotation requirements while maximizing performance on small datasets.