Preprint reports tailored LLM prompts improved detection of local health-data identifiers

A preprint benchmarking de-identification on 100 pediatric-oncology notes from Texas Children's Hospital reports that institution-specific LLM prompting identified protected health information that purpose-built systems and reference annotations missed. Its best LLM result had F1 of 0.918 versus 0.779 for Stanford TiDE, while a final prompt reached 0.981 recall against 227 newly confirmed PHI spans. The authors found that naming local identifier categories recovered 48 of 61 previously missed categories; multi-agent configurations did not outperform calibrated single-pass prompts.