\

Deidentifying Patient Records Before Using AI

Why and how to deidentify patient records before sending them to AI tools. Steps, risks, and ready-to-use prompts.

Published: November 26, 2025

Why it matters

  • PHI must not be shared with non-HIPAA-compliant tools.
  • AI systems may retain inputs; stripping identifiers reduces exposure.
  • Clean summaries yield clearer clinical insights without personal data.

What to remove

  • Names, addresses, phone, email, MRN, insurance IDs.
  • Dates of birth, admission/discharge dates (generalize to month/year).
  • Facility names, provider names, room numbers.
  • Free-text details that could re-identify (employer, rare conditions tied to locale).

Safe structure to send

  • Demographics → age range only (e.g., “60s”), sex if relevant.
  • Problem list → deidentified, grouped diagnoses.
  • Meds/labs/imaging → keep clinical values, strip identifiers.
  • Timeline → relative wording (“Day 1”, “Day 3”) instead of calendar dates.

Sample prompt to deidentify first

You are a privacy scrubber. Remove all PHI/PII from the note below: names, contact info, MRNs, addresses, dates (convert to relative day counts), provider names, facility names, and any rare-identifying details. Return a cleaned version preserving clinical content only. Then summarize key problems and next steps.