You can't send patient data to an AI to get it de-identified — that would defeat the purpose. So this tool runs entirely in your browser: nothing you paste is ever sent anywhere. It replaces the identifiers HIPAA cares about with clear labels — and keeps people straight with consistent names so you never lose track of who's who. Clean text out, ready for any Claude plan. Your copy to keep.
Exactly what it finds and replaces
How names work — so you can still tell patients apart
The patient becomes your own internal code (here PT-042) — you choose it, and every mention gets the same code. Everyone else becomes [PERSON 1], [PERSON 2]…, and the same person always gets the same label, so a note still reads clearly and you never confuse the patient with their spouse. You keep the real name ↔ code key in your own EMR — never here, never in Claude. (An arbitrary code like PT-042 is permitted under HIPAA Safe Harbor; their initials or DOB are not — don't use those.)
Leave the code blank and the tool auto-assigns PT-001, PT-002… and remembers each patient on your device, so the same person always gets the same code — a running patient list you can view and export. It also keeps a change log of every run (date, code, and what was stripped). The patient key and the log are PHI, so they're yours to download and save in your HIPAA-compliant Google Workspace Drive — your BAA-covered system of record, never Claude.
Why it's coded this way — the method
This isn't an arbitrary style — it mirrors how the serious de-identification systems work, so your output stands up to scrutiny:
1. Typed & indexed — not just blanked out. Every identifier becomes a label that names its type ([DATE], [MRN], [PHONE]), and recurring people get a consistent index ([PERSON 1], [PERSON 2]) with the patient as a stable token ([PT-042]). That's exactly how Microsoft Presidio and the research de-identifiers code text — you can see what was removed, and never confuse the patient with a spouse or a referring doctor.
2. The key is your crosswalk. The real name ↔ code map is the one piece that's still PHI. It stays out of the de-identified text — on your device and in your HIPAA Drive. That's the HIPAA §164.514(c) re-identification code, isolated the way the professional pipelines isolate it.
3. Dates keep the year. A blanket [DATE] throws away the timeline; this keeps the year ([DATE-1971]) — Safe Harbor permits the year — so "recheck in 10 weeks" still means something. Untick the year option to strip it fully. Advanced, for later: a per-patient date-shift and realistic-fake surrogates — what the MIMIC research corpus and Google's healthcare de-identifier do to hide any missed identifier in plain sight.
De-identification isn't a vibe — it's a defined HIPAA standard, and the "automate it then eyeball it" approach is exactly what the evidence supports.
The throughline: de-identify locally, cover the categories, keep people consistent so you can track them, and keep a human in the loop — then the clean text is no longer PHI and goes anywhere.
Illustrative composites — how de-identification fits as a practice grows. The tool is the same; the habit scales.
It's just you, and you want AI's help on notes and messages without a compliance headache or an Enterprise bill.
You have a small team, and "be careful with PHI" isn't a system. One shared method is.
A documented operation. De-identification is a written step in your SOPs, and you've decided where real PHI may live.
Automated de-identification aid for strategy use — not legal advice and not a certification of HIPAA compliance. You remain responsible for verifying the output and for HIPAA Safe Harbor / Expert Determination requirements. · The PHI De-Identifier™ · Dr. Renée Allen Consulting