COIOS
Emergent technology Item new preprint

A local multi-agent AI system deidentified multimodal hospital data at 99.8% sensitivity

A locally deployable multi-agent system combining multimodal large language models, task-specific networks and rule-based steps removed 99.82% of injected identifiers across text, imaging, handwriting and audio while preserving 99.61% of clinically important content.

The authors built and benchmarked a modular, on-premises multi-agent deidentification system across 16 orchestrator configurations, using 250 MIMIC-IV records with synthetically injected identifiers alongside head CT, face images, handwriting, audio and German clinical text. The best local configuration reached per-identifier sensitivity of 99.82% (95% CI 99.76-99.88) with clinical preservation of 99.61% per file, and matched a proprietary model on sensitivity with higher specificity. On the institution's own data, per-patient sensitivity was 100%.

Why it is interesting: Deidentification that runs inside the hospital and keeps clinical content intact is the bottleneck for reusing multimodal records in research.

Source
medRxiv (preprint), 2026-09-14
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
Driver
What linked data reveals