EdgeDeID de-identifies clinical notes locally without real patient training data
Original title: EdgeDeID: A 0.6B Clinical De-identifier Trained Without Real Patient Data
On a 100-note synthetic benchmark it reaches 0.994 micro-F1, but F1 falls to 0.48 on MIMIC ICU nursing notes.
Sending clinical notes to a hosted model exposes the identifiers de-identification is meant to remove. EdgeDeID is a 0.6B decoder-only model trained on synthetic notes written by three LLM teachers and relabeled into one convention. It emits a typed XML schema of identifiers and, quantized, runs on CPU in 615 MB. Only a verifier adapter among three training additions helped, separating complete from incomplete outputs at 0.94 AUROC. Three handwritten in-context examples recover 0.08 F1 on MIMIC without annotated target data.
Unconfirmed: In-domain scores use a synthetic benchmark with manually verified annotations, not real patient notes.
- It left no missed identifier in 99% of synthetic in-domain notes.
- The strongest off-the-shelf system scored 0.857 micro-F1.
- That baseline left no missed identifier in 12% of notes.
- Reinforcement learning and rule-based post-processing left in-domain results unchanged.
- Errors track telegraphic style and low identifier density absent from synthetic notes.
From the original article
1 min read at arxiv.orgDe-identifying clinical notes with a hosted language model discloses the information the step is meant to protect. We present EdgeDeID, a 0. 6B-parameter decoder-only de-identifier that runs locally and is trained without real patient text: three LLM teachers write synthetic notes, a single labeler re-annotates them into one convention, and the model learns to emit a typed XML schema of the identifiers it finds.
By Sina Daneshgar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.