Researchers have released a tool stack that lets hospitals de-identify patient records locally, sidestepping the cloud entirely.
The SHIELD dataset — Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification — contains 1,381 clinical notes with 10,229 labeled spans of protected health information across 9 categories. The team used diversity sampling to ensure notes span demographic and document-type variety that older benchmarks like i2b2 2006 and 2014 simply do not have. They then used large language models as teachers to train smaller, locally deployable models through a distillation framework. The best distilled model, based on DeBERTa v3, hits 0.89 precision and 0.88 recall on span-level detection while running on standard workstation hardware.
The real unlock here is governance, not performance. Hospital IT teams have long faced a catch-22: the most capable de-identification models are cloud-hosted, but sending records containing protected health information to external APIs invites regulatory and legal risk. A model that runs behind institutional firewalls and still approaches cloud-teacher accuracy removes that tradeoff. The dataset and model are both publicly released.
The system is not gap-free — institution-specific entity types transfer poorly across settings, and per-category recall still lags the cloud teacher (0.81 vs. 0.90 macro-averaged), which means high-stakes deployments will likely need a specialized supplemental model for edge categories rather than a single off-the-shelf solution.