A locally deployable agentic framework for clinical data deidentification

The Multimodal Anonymizer achieves near-complete deidentification of multimodal hospital data with over 98% sensitivity and preserves critical clinical information at rates above 99%.

• Why it matters: Safe reuse of hospital data for AI development is hindered by unreliable, modality-specific deidentification methods that risk removing important clinical content, limiting research progress.
• What they did: A modular, fully local multi-agent framework integrating multimodal large language models, neural networks, and rule-based transformations was developed and evaluated across 16 configurations using diverse datasets, including hospital and public data.
• The result: The best configuration matched proprietary models in sensitivity, outperformed existing tools across modalities, and enables safer large-scale reuse of clinical data while maintaining high clinical content preservation.