The Multimodal Anonymizer: a fully local multi-agent AI system for medical data deidentification
- 1Charite Universitaetsmedizin Berlin, Institute for Artificial Intelligence in Medicine, Invalidenstrasse 120, 10115 Berlin, Germany
Abstract
Multimodal clinical data contain identifiers across diverse formats. We developed the Multimodal Anonymizer, a locally deployable framework combining multimodal language models, specialist networks, deterministic transformations and iterative verification. We compared 16 orchestrator configurations using automated evaluation on 250 constructed MIMIC-IV-derived patient bundles with synthetic identifiers. A selected local configuration underwent manual evaluation on these same bundles supplemented with head CT scans, face images, handwritten notes, audio recordings and synthetic German clinical notes from unrelated sources. Separate evaluations used the German GRASCCO corpus and 250 Charite partograms. On this expanded benchmark, patient-level deidentification sensitivity was 98.80% (95% CI, 97.20-100), per-identifier sensitivity 99.82% (95% CI, 99.76-99.88), patient-level critical clinical preservation 99.60% (95% CI, 98.80-100) and the clinical preservation score 99.61% (95% CI, 99.51-99.71). All partograms met the study-defined deidentification criterion. These findings support feasibility, while constructed bundles, synthetic identifiers, limited natural-data validation and absent component ablations constrain generalizability.
Related
- Medical students' appraisal of curated AI answers in a cardiovascular system course
- Medical image encryption algorithm based on Chen’s chaotic system and the Advanced Encryption Standard algorithm
- The VOICE-DEP study protocol: Multimodal analysis of voice and speech during medical interviews to support diagnosis and longitudinal monitoring of major depressive disorder
- Student Characteristics Related to the Usability and Tolerability of a Metaverse-Based Anatomy Education System: A Cross-Sectional Study of 64 Medical Students
- Autonomic nervous system dysfunction in critically ill cancer patients: Hemodynamic consequences and multimodal assessment strategies☆
BibTeX
@misc{hirsch2026multimodal,
title = {{The Multimodal Anonymizer: a fully local multi-agent AI system for medical data deidentification}},
author = {A. Hirsch and F. W. Ten and K. S. Krueger and R. Geyer and T. Roeschl and M. Groeschel and P. Rostin and R. Eils and M. Spott and F. Prasser and A. Meyer and J. Madrid},
howpublished = {medRxiv},
year = {2026},
month = may,
doi = {10.64898/2026.05.28.26353952},
url = {https://www.medrxiv.org/content/10.64898/2026.05.28.26353952v2},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - The Multimodal Anonymizer: a fully local multi-agent AI system for medical data deidentification
AU - A. Hirsch
AU - F. W. Ten
AU - K. S. Krueger
AU - R. Geyer
AU - T. Roeschl
AU - M. Groeschel
AU - P. Rostin
AU - R. Eils
AU - M. Spott
AU - F. Prasser
AU - A. Meyer
AU - J. Madrid
PY - 2026
DA - 2026/05/28
PB - medRxiv
DO - 10.64898/2026.05.28.26353952
UR - https://www.medrxiv.org/content/10.64898/2026.05.28.26353952v2
AB - Multimodal clinical data contain identifiers across diverse formats. We developed the Multimodal Anonymizer, a locally deployable framework combining multimodal language models, specialist networks, deterministic transformations and iterative verification. We compared 16 orchestrator configurations using automated evaluation on 250 constructed MIMIC-IV-derived patient bundles with synthetic identifiers. A selected local configuration underwent manual evaluation on these same bundles supplemented with head CT scans, face images, handwritten notes, audio recordings and synthetic German clinical notes from unrelated sources. Separate evaluations used the German GRASCCO corpus and 250 Charite partograms. On this expanded benchmark, patient-level deidentification sensitivity was 98.80% (95% CI, 97.20-100), per-identifier sensitivity 99.82% (95% CI, 99.76-99.88), patient-level critical clinical preservation 99.60% (95% CI, 98.80-100) and the clinical preservation score 99.61% (95% CI, 99.51-99.71). All partograms met the study-defined deidentification criterion. These findings support feasibility, while constructed bundles, synthetic identifiers, limited natural-data validation and absent component ablations constrain generalizability.
ER -