Contextualized medication information extraction using Transformer-based deep learning architectures.

Area of Science:

• Natural Language Processing (NLP)
• Clinical Informatics
• Artificial Intelligence in Healthcare

Background:

• Extracting medication information and its context from clinical narratives is crucial for understanding patient treatment.
• Previous methods often struggle with the complexity and nuances of clinical text.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) system for extracting medications and contextual information related to drug changes.
• To assess the performance of state-of-the-art transformer models, including the large language model GatorTron, for these tasks.

Main Methods:

• Developed NLP systems for medication mention extraction, event classification, and context classification.
• Explored six pretrained transformer models, with a focus on GatorTron, pretrained on extensive clinical text.
• Evaluated system performance using annotated data and scripts from the 2022 n2c2 challenge.

Main Results:

• GatorTron achieved top rankings: 3rd for medication extraction (F1-score 0.9828), 2nd for event classification (F1-score 0.9379), and best micro-average accuracy for context classification (0.9126).
• GatorTron outperformed other transformer models, demonstrating the benefits of large-scale pretraining on clinical data.

Conclusions:

• Large transformer models, particularly GatorTron, offer significant advantages for extracting contextual medication information from clinical narratives.
• This NLP approach enhances the ability to understand drug changes within electronic health records.