Text Mining of Adverse Events in Clinical Trials: Deep Learning Approach.

Area of Science:

• Pharmacovigilance and Drug Safety
• Clinical Trial Monitoring
• Medical Informatics

Background:

• Pharmacovigilance is crucial for identifying unknown adverse events and pattern changes during medicine use.
• Effective safety reporting is essential for drug development and patient well-being.

Purpose of the Study:

• To demonstrate the feasibility of automating the coding of adverse events from narrative text in serious adverse event reports.
• To enable statistical analysis of adverse event patterns for timely correlation with trial medications.

Main Methods:

• Utilized the Unified Medical Language System (UMLS) for coding, integrating multiple terminologies.
• Employed MetaMap for identifying UMLS concepts within report narratives.
• Developed a Bidirectional Encoder Representations from Transformers (BERT) binary classifier to distinguish adverse events from other mentions.

Main Results:

• The BERT model achieved a high F1 score of 0.8080, successfully handling class imbalance.
• Performance was 10.15% lower than human performance but 17.45% higher than baseline methods.
• Demonstrated the model's capability in accurately coding adverse events from unstructured text.

Conclusions:

• Automated coding of adverse events in serious adverse event reports is feasible.
• Statistical analysis of coded adverse events can reveal timely correlations with investigational medicines.
• This approach enhances the efficiency of drug safety monitoring in clinical trials.