generating-synthetic-surrogates

Generating synthetic surrogates

method="replace" swaps each detected identifier for a realistic, type-matched fake — John Doe becomes Mark Lee, a phone becomes a plausible phone, a date becomes a plausible date. Unlike opaque [REDACTED]/[NAME] masks, surrogate text reads naturally and stays parseable by downstream NLP, while still containing no real PHI. OpenMed generates surrogates on-device via Faker-backed providers keyed to each canonical label.

When to use this skill

Use surrogates when the de-identified text must remain readable or machine- parseable: training data for clinical NLP, demos, QA, or notes a human still needs to skim. If you only need the identifiers gone and don't care about readability, plain method="mask" is simpler and more obviously redacted.

Quick start

import openmed

note = (
    "Patient John Doe (MRN 1234567) saw Dr. John Doe's colleague on 2024-03-02. "
    "Reach John Doe at 617-555-0142."
)

result = openmed.deidentify(
    note,
    method="replace",
    consistent=True,     # every "John Doe" -> the SAME surrogate within this call
    seed=42,             # reproducible across runs
    locale="en_US",      # shapes the fakes; defaults from lang via LANG_TO_LOCALE
)
print(result.deidentified_text)
# Patient Mark Lee (MRN 8830127) saw Dr. Mark Lee's colleague on 2024-07-18. ...

consistent=True is what makes the output coherent: the three mentions of "John Doe" collapse to one fake identity instead of three different ones, so the note still makes sense. seed= makes that mapping reproducible run to run.

Surrogates vs opaque redaction