{"slug": "edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data", "title": "EdgeDeID: A 0.6B Clinical De-identifier Trained Without Real Patient Data", "summary": "A 0.6B-parameter decoder-only de-identifier called EdgeDeID, trained entirely on synthetic clinical notes written by three LLM teachers and re-annotated by a single labeler, reached 0.994 micro-F1 on a 100-note synthetic in-domain benchmark with manually verified annotations, leaving no missed identifier in 99% of notes, versus 0.857 and 12% for the strongest of eight off-the-shelf systems, according to the arXiv paper 2610.33839v1. Quantized, EdgeDeID runs locally on CPU in 615 MB, avoiding the disclosure risk of sending clinical notes to a hosted model. On MIMIC ICU nursing notes F1 falls to 0.48 because of telegraphic style and low identifier density absent from the synthetic data, and three hand-written in-context examples recover 0.08 F1 without annotated target data; of three additions to supervised training, only a verifier adapter helped, separating complete from incomplete outputs of other systems at 0.94 AUROC.", "body_md": "arXiv:2610.33839v1 Announce Type: new \nAbstract: De-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. On a 100-note synthetic in-domain benchmark with manually verified annotations it reaches 0.994 micro-F1 and leaves no missed identifier in 99% of notes, against 0.857 and 12% for the strongest of eight off-the-shelf systems. Quantized, it runs on CPU in 615 MB. Of three additions to supervised training, only a verifier adapter helps: it separates complete from incomplete outputs of other systems at 0.94 AUROC, whereas reinforcement learning with a recall-weighted reward and a rule-based post-processing layer leave in-domain results unchanged. On MIMIC ICU nursing notes F1 falls to 0.48. The errors follow two properties our synthetic notes lack, telegraphic style and low identifier density, and three hand-written in-context examples recover 0.08 F1 without annotated target data.", "url": "https://wpnews.pro/news/edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data", "canonical_source": "https://www.machinebrief.com/news/edgedeid-a-06b-clinical-de-identifier-trained-without-real-p-rryy", "published_at": "2026-10-27 04:00:00+00:00", "updated_at": "2026-10-27 05:42:11.224041+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-safety", "natural-language-processing"], "entities": ["EdgeDeID", "arXiv", "MIMIC"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data", "markdown": "https://wpnews.pro/news/edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data.md", "text": "https://wpnews.pro/news/edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data.txt", "jsonld": "https://wpnews.pro/news/edgedeid-a-0-6b-clinical-de-identifier-trained-without-real-patient-data.jsonld"}}