Contextualized medication event extraction with striding NER and multi-turn QA.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Text Mining

Background:

• Automated extraction of medication events from clinical notes is crucial for patient safety and pharmacovigilance.
• Identifying medication changes and their associated contexts (e.g., reasons, timing) remains a challenge in clinical text analysis.

Purpose of the Study:

• To develop and evaluate a system for contextualized medication event extraction from clinical notes.
• To accurately identify medication name spans, classify medication change events, and determine their contexts.

Main Methods:

• A pipeline system combining a striding Named Entity Recognition (NER) model for medication extraction with ensemble classification models (span-based and question-answering) for event and context classification.
• The striding NER model processes clinical text using overlapping subsequences and large pre-trained language models.
• Multi-turn question-answering and span-based models were employed for event and context classification.

Main Results:

• The system achieved a combined F-score of 66.47% for end-to-end contextualized medication event extraction on the n2c2 2022 Track 1 dataset.
• This performance represents the highest score among participants in the n2c2 2022 Track 1 challenge.
• The system demonstrated effectiveness in medication extraction (ME), event classification (EC), and context classification (CC).

Conclusions:

• The developed system effectively extracts contextualized medication events from clinical notes.
• The proposed pipeline architecture, integrating striding NER with ensemble classification, is highly effective for clinical text analysis.
• This work advances the state-of-the-art in automated medication event extraction from electronic health records.