Hybrid Deep Learning for Medication-Related Information Extraction From Clinical Texts in French: MedExt Algorithm...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Patient medication information is vital for healthcare but largely exists in unstructured text (up to 80%).
• Manual extraction of medication data from French clinical text is challenging and under-researched.
• Developing automated methods for extracting medication information from French corpora is needed.

Purpose of the Study:

• To create an automated system for extracting medication-related information from French clinical text.
• To address the gap in natural language processing research for French medical corpora.
• To improve the accessibility and usability of patient medication data.

Main Methods:

• A hybrid system was developed, integrating expert rules, contextual word embeddings, and a deep recurrent neural network (bidirectional long short term memory-conditional random field).
• The system was trained and evaluated on 320 manually annotated French clinical notes.
• Performance was compared against rule-based and machine learning-only approaches using token-level recall, precision, and F-measure.

Main Results:

• The hybrid system achieved an overall F-measure of 89.9% (90.8% precision, 89.2% recall).
• Combining expert rules and contextualized embeddings significantly improved performance compared to methods without these components.
• High F-measures were obtained for medication name (95.3%) and dosage (95.3%), with notable performance in frequency (92.2%).

Conclusions:

• Integrating expert rules with deep contextualized embeddings and neural networks enhances medication information extraction.
• A synergistic effect was observed between expert knowledge and latent knowledge derived from data.
• The developed system offers a promising solution for extracting critical medication data from French clinical text.