New AI model de-identifies clinical notes using synthetic data

Researchers have developed EdgeDeID, a small, 0.6 billion parameter language model designed for clinical de-identification that operates locally. This model was trained entirely on synthetic data generated by other LLMs, avoiding the use of real patient information. EdgeDeID achieves a high micro-F1 score of 0.994 on a synthetic benchmark and can run on a CPU with minimal memory. While effective on synthetic data, its performance drops significantly on real-world clinical notes from MIMIC ICU, indicating a gap in handling real-world data characteristics. AI

IMPACT This model offers a privacy-preserving approach to clinical data de-identification, potentially enabling more secure handling of sensitive health information.

RANK_REASON The cluster describes a research paper detailing a new model and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• EdgeDeID
• MIMIC ICU

• paper
• model release
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →