PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation

Researchers developed PLACID, a privacy-preserving system using small on-device AI models (2B-10B parameters) for clinical acronym disambiguation in healthcare settings. The cascaded approach combines general-purpose models for detection with domain-specific biomedical models, achieving 81% expansion accuracy while keeping sensitive health data local.

• →Small on-device AI models can provide clinical acronym disambiguation without sending sensitive health data to external servers.
• →General instruction-following models achieve high detection accuracy (~98.8%) but poor expansion capabilities (~65.5%) for medical acronyms.
• →A cascaded pipeline using domain-specific medical models improves expansion accuracy to approximately 81%.
• →Privacy-preserving AI solutions can address healthcare's strict data protection requirements while maintaining functionality.
• →On-device models with 2B-10B parameters demonstrate viable alternatives to cloud-dependent large language models for specialized healthcare tasks.