Detecting Adverse Drug Events in Clinical Notes Using Large Language Models.

Area of Science:

• Pharmacovigilance and Patient Safety
• Medical Informatics
• Natural Language Processing in Healthcare

Background:

• Monitoring adverse drug events (ADEs) is crucial for patient safety and pharmacovigilance.
• Identifying ADEs is challenging due to their frequent documentation in unstructured clinical notes within electronic health records (EHRs).
• Manual review of clinical notes for ADE detection is inefficient and time-consuming, necessitating automated solutions.

Purpose of the Study:

• To fine-tune and evaluate a large language model (LLM) for the automated detection of ADEs in clinical notes.
• To assess the feasibility of using LLMs for extracting ADE-related information from unstructured EHR data.
• To highlight the challenges and potential improvements in ADE reporting practices.

Main Methods:

• Fine-tuning a large language model (LLM) specifically for the task of ADE detection.
• Evaluating the LLM's performance on a dataset of clinical notes.
• Descriptive analysis of ADE documentation within discharge notes.

Main Results:

• Preliminary results indicate that ADEs are infrequently and poorly documented in discharge notes.
• Less than 15% of documented ADEs were explicitly linked to specific drugs in the analyzed notes.
• This highlights significant gaps in current ADE reporting within EHRs.

Conclusions:

• Automated methods, such as fine-tuned LLMs, show promise for improving ADE detection in clinical documentation.
• Current ADE reporting practices in EHRs require significant improvement to ensure accurate and comprehensive pharmacovigilance.
• Further research is needed to enhance the accuracy and completeness of ADE identification and reporting.