Efficient extraction of medication information from clinical notes: an evaluation in 2 languages

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Text Mining

Background:

• Extracting medication information from clinical narratives is crucial for patient safety and pharmacovigilance.
• Existing natural language processing (NLP) methods may have limitations in accuracy, computational cost, or portability across languages.

Purpose of the Study:

• To evaluate a novel transformer-based NLP architecture for extracting medication details from clinical text.
• To assess the accuracy, computational efficiency, and cross-lingual portability of the proposed method.

Main Methods:

• Developed and trained a transformer-based architecture for named entity recognition and relation extraction of medication information.
• Evaluated the model on a French clinical corpus and an English clinical corpus from the 2018 n2c2 shared task.
• Compared performance and computational cost against an existing transformer-based method.

Main Results:

• The proposed architecture achieved competitive relation extraction performance (F-measures 0.82 for French, 0.96 for English).
• End-to-end extraction (NER and RE) yielded F1 scores of 0.69 (French) and 0.82 (English).
• Significantly reduced computational cost by 10% compared to the existing method.

Conclusions:

• The novel NLP architecture demonstrates high performance in extracting medication information from both French and English clinical texts.
• The method offers a favorable balance of accuracy and reduced computational impact, suitable for hospital IT environments.
• This approach enhances the portability and efficiency of medication information extraction in diverse clinical settings.