[TAGALOG]
The 50% Miss Rate Problem
A 2025 survey of LLM-based de-identification tools (arXiv:2509.14464) found that general-purpose LLM tools miss more than 50% of clinical PHI in multilingual documents. This figure reflects a fundamental architectural mismatch: LLMs are designed for language understanding and generation, not for the structured, high-recall identification task that HIPAA de-identification requires.
The HIPAA Privacy Rule's Safe Harbor method requires removal of 18 specific iden...