New SHIELD Dataset Addresses Anonymisation of Clinical Notes

Vad har hänt

Researchers have developed SHIELD (Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification), a dataset of 1,394 clinical notes containing 10,505 Protected Health Information (PHI) annotations across nine categories. The dataset, created using diversity sampling and human review, aims to modernise existing anonymisation benchmarks. SHIELD is used to evaluate the performance of four large language models (LLMs) and distill their capabilities into locally deployable small language models (SLMs).

Varför det spelar roll

The anonymisation of clinical texts is essential for the secondary use of electronic health records (EHR) in research and development without compromising patient privacy. Existing public benchmarks, such as i2b2 2006/2014, are over a decade old and lack the semantic and demographic diversity found in modern records. The new dataset and models address the need for more relevant tools for anonymising sensitive data. While large language models are effective, they involve high computational costs and can pose data governance challenges when PHI is processed via cloud-based APIs. Distillation into small language models enables more efficient local deployment, better complying with PHI regulations.

Vem påverkas

NLP and AI researchers, healthcare professionals, hospital administrators, and software developers working with Electronic Health Records (EHR) are affected. Organisations handling patient records can benefit from improved tools to securely anonymise data, thereby enabling secondary data use for research and improved care. Patients and their privacy are the ultimate stakeholders, as robust anonymisation protects sensitive information.

EU-status

Ej relevant för EU-status.

Mer att veta

The authors highlight that the SHIELD data is based on synthetic, human-annotated identifier-replaced entries. This distinguishes it from existing datasets that exclusively use real patient data.