Contextualized medication event extraction with levitated markers.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Artificial Intelligence

Background:

• Automatic extraction of patient medication histories from clinical notes is crucial for treatment planning.
• Clinical text mining requires identifying not only medication events but also their context (negation, uncertainty, timing).
• Accurate patient timelines depend on comprehensive contextual information from clinical notes.

Purpose of the Study:

• To introduce Levitated Context Markers (LCMs), a novel transformer-based model for contextualized event extraction.
• To adapt levitated markers for pretrained transformer models to use global input representations and focus on event-related subspans.
• To improve the accuracy of extracting medication events and their context from free-text clinical notes.

Main Methods:

• Developed Levitated Context Markers (LCMs), a transformer-based model.
• Adapted levitated markers using a sparse attention mechanism for focused event-related subspan analysis.
• Utilized pretrained transformer models to process global input representations.

Main Results:

• LCMs outperformed a strong baseline model on the Contextualized Medication Event Dataset.
• Demonstrated that LCMs' sparse attention provides interpretable predictions.
• Showcased the ability to detect relevant context cues in an unsupervised manner.

Conclusions:

• Levitated Context Markers (LCMs) represent a significant advancement in contextualized event extraction for clinical text.
• The model enhances the extraction of patient medication histories by accurately capturing event context.
• LCMs offer interpretable insights into context detection, aiding in the construction of precise patient timelines.