Detecting Adverse Drug Events with Rapidly Trained Classification Models.

Area of Science:

• Computational linguistics
• Medical informatics
• Clinical NLP

Background:

• Identifying medication side effects and adverse drug events (ADEs) from clinical narratives is crucial but challenging.
• Formal reporting of ADEs often misses information present in unstructured clinical notes.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) system for identifying mentions of symptoms, drugs, and their relationships (indications or ADEs) in clinical notes.
• To improve the detection of ADEs using innovative feature engineering.

Main Methods:

• Utilized a natural language processing (NLP) system incorporating word embeddings with dimensionality reduction.
• Employed a conditional random field (CRF) model for named entity recognition (NER) and a random forest model for relation extraction (RE).
• Evaluated system performance on a manually annotated dataset and submitted to the MADE 1.0 competition.

Main Results:

• Achieved micro-averaged F1 scores of 80.9% for NER and 88.1% for RE.
• The integrated system achieved an F1 score of 61.2%.
• The relation extraction system ranked first in Task 2, and the integrated system ranked third in Task 3 of the MADE 1.0 challenge.

Conclusions:

• The study demonstrates the effectiveness of NLP in detecting ADEs from electronic health records.
• Innovative feature engineering significantly benefits the performance of NLP systems for ADE identification.