Hospitals Protect the Image but Leak the Report: Preprint Benchmarks Pseudonymisation with 15 Detectors, 4 Languages
maier_ak · x · 2026-10-12
Researchers from FAU, Universitätsklinikum Erlangen and Imperial College London released the preprint "The Text Beside the Image", measuring detection, utility and re-identification leakage of clinical text pseudonymisation on the same documents for the first time, across 15 detectors (rule engine, German domain tagger, two zero-shot models, multilingual NER, two clinical de-identifiers, eight LLMs), 4 corpora and 4 languages.
Key findings:
- Hospital release efforts focus on the image, but the accompanying report or discharge letter is where the name lives — a carefully protected scan doesn't compensate for a report naming the patient.
- Combination strategy matters: requiring two detectors to agree drops sensitivity to 0.83, raises linkage risk fivefold, and lets clear-text names jump from 152 to 1,816. Spend on recall, not cleverer surrogates.
- On German discharge letters, a 13-detector union reaches 0.9998 person sensitivity: one surviving mention in 2,800, one letter in 400, with all attacks landing at chance — with a large enough ensemble, pseudonymising clinical free text has become very safe.
More from Safety
- Education as an AI safety area: Ruxandra Teslo argues classrooms are an alignment frontline — mikeflache · 2026-10-12
- Caine Prize Winner Flagged as Almost Entirely AI-Generated by Pangram — ZeroStateReflex · 2026-10-12
- TechnoLama shows the AI-recreated PowerPoint lookalike from the ArtCraft suite — technollama · 2026-10-12
- From nuclear safety to AI: containing agents with signals they can't produce — WeldPond · 2026-10-12
- Paying via AI agents is currently illegal in Europe, warns exec — emmanuelvivier · 2026-10-12
- Security researcher calls Anthropic's Cyber Verification 'absolutely useless' — matthew_d_green · 2026-10-12