Developer Finds Rival Clinical De-Identification Project Outperforms Their Own After Post-Processing Fixes

A developer building localscrub, a local-first tool for scrubbing identifiers from clinical notes, benchmarked it against OpenMed, a larger Apache-2.0 healthcare AI project with thousands of stars. Running OpenMed's default PII extraction on 100 public transcripts with 697 planted identifiers produced precision of only 0.62 and a name F1 of 0.00, despite near-perfect recall. The author traced the weakness to three post-processing issues: a SWIFT/BIC regex mislabeling section headers as identifiers, an overly broad Safe Harbor policy that masked non-required details, and fragmented spans that left parts of names and addresses exposed. Once these were corrected, precision rose to 0.91 and name F1 to 0.96, which prompted the author to weigh whether to publish a comparison or pursue a different path.