Privacy-preserving mimic models for clinical named entity recognition in French

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Clinical narrative notes contain vital patient information but are unstructured.
• Deep learning for clinical Named Entity Recognition (NER) requires extensive annotated data.
• Limited public medical corpora exist due to data sensitivity.

Purpose of the Study:

• To develop privacy-preserving, shareable models for French clinical NER.
• To enable knowledge transfer from a private teacher model to a publicly shareable student model using mimic learning.
• To address the scarcity of annotated medical data for NER tasks.

Main Methods:

• Implemented mimic learning for knowledge transfer between a teacher and student model.
• Trained models on a private corpus and evaluated privacy-preserving student models on three French medical corpora.
• Compared performance against baseline dictionary-based models.

Main Results:

• A student model trained with silver annotations achieved a macro F-measure of 70.6%.
• The original private teacher model achieved 85.7% F-measure.
• Privacy-preserving mimic learning models demonstrated a balance between performance and data privacy.

Conclusions:

• Mimic learning effectively transfers knowledge while preserving data privacy in clinical NER.
• The developed models offer a viable solution for sharing clinical NER capabilities without compromising sensitive patient information.
• This approach facilitates the creation of valuable, shareable resources for medical NLP research.