Perspectives in clinical NLP - Privacy and Temporality

The first part of this presentation is a hybrid approach to automatically remove Personal Health Identifiers (PHIs) from clinical text using medicine reconciliation documents from a UK hospital. Three models—Conditional Random Fields (CRF), BERT, and MedCAT—were integrated and evaluated on five PHI categories: names, dates, locations, contacts, and ID numbers. The hybrid approach combining CRFD and BERT achieved the best overall performance (F1: 0.78), with precision reaching 0.96 for BERT and MedCAT combinations. Results demonstrate that optimal model combinations vary by entity type, suggesting the importance of customizable de-identification pipelines. Key challenges included recognizing local institutional codes and handling incomplete temporal expressions.

The second of this presentation tackles temporal relation extraction from French rare genetic disease narratives, where diagnostic delays average 5-7 years across over 7,000 diseases. A comprehensive annotation framework was developed encompassing six temporal entity types (DATE with 6 subtypes, TIME, AGE, DURATION, FREQUENCY, EVENT) and eight temporal relation types linking phenotypes to timepoints. The annotated corpus comprises 490 clinical reports with 12,464 entity-relation pairs. Systematic evaluation of multiple extraction approaches—Parameter-Efficient Fine-Tuning (PEFT) with CamemBERT-Bio, OpenNRE toolkit, and large language models (LLaMA 8B, Mistral-small 22B)—revealed that PEFT achieved superior F1 scores across most relation types while maintaining computational efficiency suitable for privacy-preserving local deployment in hospital settings.

Slides