Extracting Adverse Drug Event Information with Minimal Engineering.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Information extraction (IE) is crucial for identifying drug-related information from clinical text.
• Classical machine learning methods have been extensively researched for feature engineering in IE.
• Emerging neural network approaches offer potential for improved performance in complex NLP tasks.

Purpose of the Study:

• To evaluate and compare the efficacy of classical information extraction methods against contemporary neural methods for extracting drug-related attributes, including adverse drug events.
• To benchmark performance using the 2018 N2C2 shared task dataset.

Main Methods:

• Support Vector Machine (SVM) classifiers were trained for detecting drug and drug attribute spans.
• Detected entities were paired as training instances for an SVM relation classifier, utilizing standard features.
• Baseline neural methods employing contextualized embedding representations were used for comparison in both entity and relation extraction.

Main Results:

• Both SVM-based and neural systems achieved comparable overall results.
• The SVM system demonstrated superior performance in concept extraction.
• The neural system outperformed the SVM system in relation extraction tasks.

Conclusions:

• Classical and neural IE methods yield comparable results for drug-related attribute extraction.
• Neural networks show a surprising advantage in relation extraction, despite the extensive research in classical IE feature development.