Adverse drug event detection using natural language processing: A scoping review of supervised learning methods.

Area of Science:

• Pharmacovigilance
• Health Informatics
• Computational Linguistics

Background:

• Adverse drug events (ADEs) pose significant risks in hospitals, necessitating robust monitoring systems.
• Natural language processing (NLP) offers a potential solution for automated ADE detection in clinical text.
• A critical appraisal of NLP methods for hospital-based ADE monitoring is currently lacking.

Purpose of the Study:

• To conduct a scoping review of NLP methods for ADE detection in electronic health records (EHRs).
• To critically appraise the performance and application of NLP techniques in hospital ADE monitoring.
• To identify knowledge gaps and provide directions for future research and practice.

Main Methods:

• A systematic scoping review was performed, screening 1,065 articles for eligibility.
• Included studies focused on NLP application for ADE detection in inpatient clinical narratives within EHRs.
• Extracted data included NLP methods, tasks (e.g., named entity recognition, relation extraction), and performance metrics.

Main Results:

• Twenty-nine articles met the inclusion criteria, with named entity recognition and relation extraction being the most common NLP tasks.
• Long Short-Term Memory and Conditional Random Field models were frequently employed.
• Reported high overall performance may be inflated due to significant performance drops in specific ADE-related predictions.

Conclusions:

• Treating ADEs as relations between drug and non-drug entities is recommended for corpus annotation.
• Future research should explore semi-automated methods to reduce annotation burden and investigate practical implementation of NLP in clinical settings.
• Improved NLP approaches are crucial for accurate and scalable ADE monitoring in hospitals.