Identifying Opioid Overdose and Opioid Use Disorder and Related Information from Clinical Narratives Using Large...

Area of Science:

• Public Health
• Medical Informatics
• Natural Language Processing

Background:

• Opioid overdose and opioid use disorder (OUD) represent a significant and escalating public health crisis in the U.S.
• Accurate identification of OUD-related information from clinical data is essential for research and intervention development.

Purpose of the Study:

• To evaluate and compare the effectiveness of encoder-based and decoder-based large language models (LLMs) for extracting opioid overdose and OUD information from clinical narratives.
• To identify key concepts related to opioid overdose and OUD, including problematic opioid use.

Main Methods:

• Comparison of encoder-based LLMs and decoder-based generative LLMs.
• Utilized a cost-effective p-tuning algorithm for model optimization.
• Focused on extracting nine crucial concepts related to opioid overdose and OUD.

Main Results:

• The decoder-based generative LLM, GatorTronGPT, achieved superior performance.
• GatorTronGPT obtained the best strict F1-score of 0.8637 and lenient F1-score of 0.9057.
• Demonstrated the efficiency of generative LLMs in extracting opioid-related information.

Conclusions:

• Generative LLMs, specifically GatorTronGPT, are efficient tools for extracting critical opioid overdose and OUD information from clinical narratives.
• This study provides a systematic method to facilitate opioid-related research using clinical text data.