De-identifying clinical texts using biomed-clinical BERTs and comprehensive risk assessment

Abstract

This work focuses on the de-identification and risk assessment of clinical text using machine learning, emphasising the Transformer-based biomedical models \textbf{BioBERT} and \textbf{ClinicalBERT}.

The primary objective is to accurately \textit{identify} and \textit{mask} sensitive entities while maintaining data utility for research.

We conduct a detailed \textit{comparative} analysis between BioBERT, ClinicalBERT, and other models, including NER-BERT, NER-CRF, and LSTMs.

We integrate BioBERT and ClinicalBERT into the \textbf{MASK} framework to facilitate effective de-identification (\textbf{MASK-BioBERT / ClinicalBERT}). We introduce a \textbf{risk assessment} tool that evaluates the potential re-identification risks of clinical data post-de-identification. Results from our experiments evaluated on a dataset of clinical records demonstrate effectiveness of BioBERT and ClinicalBERT, revealing their state-of-the-art performances in precision and recall across various entity types. The domain-specific BERT models, especially ClinicalBERT, also show strong overall macro precision and recall scores, making them a valuable tool for de-identification in healthcare settings. Additionally, the study underscores the need to continuously improve de-identification models through finetuning and expanded real-world feedback. We also advocate for more extensive datasets and enhanced training methods to close the performance gaps in handling \textit{sensitive} information. The developed methods contribute to safeguarding patient \textit{privacy} while supporting the usability of healthcare data for medical research.