Finding medication doses in the liteature

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Pharmacology

Background:

• Medication doses are critical for drug safety and efficacy.
• Dose information is primarily available as free text in scientific literature.
• Existing systems for clinical text may not effectively extract doses from research articles.

Purpose of the Study:

• To evaluate existing drug prescription information extraction systems on scientific literature.
• To investigate the efficacy of sequence-to-sequence neural networks for dose extraction.
• To identify challenges in extracting medication doses from scientific texts.

Main Methods:

• Developed a dataset of 694 PubMed Central documents annotated for drug dose information using the i2b2 schema.
• Compared performance of existing systems with novel sequence-to-sequence models.
• Analyzed differences in scope and formatting between clinical and literature dose information.

Main Results:

• Less than half of all drug doses were found in MEDLINE/PubMed abstracts; full-text access is essential.
• New approaches are needed due to distinct characteristics of dose information in literature.
• Achieved 83.9% recall, 87.2% precision, and 85.5% F1 score for complete drug prescription extraction.

Conclusions:

• Scientific literature contains significant drug dose information not present in abstracts.
• Sequence-to-sequence models show promise for improving drug dose extraction from research articles.
• Tailored methods are required to accurately capture medication dosage from diverse scientific texts.