Automatically Recognizing Medication and Adverse Event Information From Food and Drug Administration's Adverse Event...

Area of Science:

• Pharmacovigilance
• Natural Language Processing
• Biomedical Informatics

Background:

• The Food and Drug Administration's (FDA) Adverse Event Reporting System (FAERS) contains valuable unstructured narrative data on adverse drug events (ADEs).
• Extracting critical ADE information from these narratives is challenging but crucial for identifying unknown drug toxicities.
• Current FAERS data lacks comprehensive structured information on ADE severity, causality, and detailed descriptions.

Purpose of the Study:

• To develop an annotated corpus of FAERS narratives for adverse drug events.
• To create a biomedical named entity recognition (NER) system for automated extraction of ADE-related information.
• To enhance the utility of FAERS data for post-marketing pharmacovigilance.

Main Methods:

• Developed annotation guidelines for medication and adverse event entities.
• Annotated 122 FAERS narratives (approx. 23,000 tokens) for medication and adverse event entities.
• Built and evaluated a supervised machine learning named entity tagger using diverse features.

Main Results:

• Achieved high inter-annotator agreement (over 0.9 Cohen's kappa) for annotated entities.
• The best performing named entity tagger demonstrated a 0.73 F1 score for detecting medication, adverse event, and other entities.
• Successfully developed an annotated corpus and machine learning models for ADE information extraction.

Conclusions:

• An annotated corpus and machine learning models were developed to automatically extract medication and adverse event information from FAERS narratives.
• This work represents a significant advancement in enriching FAERS data for improved post-marketing pharmacovigilance.
• Automated extraction of ADE information from FAERS narratives can facilitate timely identification of drug toxicities.