COIOSDeep risk insights through multi-disciplinary research across the globe
a close up of a piece of white paper
Item new Preprint · not peer-reviewed Added

A locally run AI system removed 99.8% of identifiers planted in patient records

Of identifiers planted in 250 MIMIC-IV records extended with scans, handwriting and audio, 99.82% were removed, with every identifier removed for 98.80% of patients.

The authors built a deidentification system that runs on hospital hardware alone, and compared 16 versions of the model that coordinates its steps. Across 250 assembled MIMIC-IV records supplemented with CT scans, face images, handwriting, audio and German notes, 99.82% of planted identifiers were removed. Every identifier was removed for 98.80% of patients, and critical clinical details were preserved for 99.60%, although the identifiers were synthetic.

Why it is interesting: A deidentification tool needing no external service, tested against planted identifiers rather than real ones.

Source
medRxiv, 14 September 2026
Paper
A locally deployable agentic framework for clinical data deidentification
Authors
Hirsch A, Ten FW, Krüger KS, Geyer RC, Röschl T, Gröschel MI et al.
DOI
10.64898/2026.05.28.26353952
Type
Preprint
Design
benchmark evaluation of 16 orchestrator model configurations on 250 MIMIC-IV patients with injected identifiers plus imaging, audio, handwriting and German clinical text, including local hospital data
Verdict
New finding
Driver
AI in medicine
Attention
4 citations (Crossref) · as of 2026-11-18

Other drivers in Data, models and systems