Free · runs locally in your browser

Strip the PHI before you paste a note into ChatGPT

Aclera Scrubs is a PHI scrubber that runs locally in your browser. Paste a note, check what it flagged, copy the cleaned text. Nothing goes to a server.

The scrubber didn't start. Reload the page, or open it in a current Chrome, Edge, Safari or Firefox.

What it does

Removes by default

  • Patient names, family names and clinician names
  • Dates of birth and any date with a month or day in it
  • Medical record numbers, account numbers, insurance and member IDs
  • Phone and pager numbers, email addresses
  • Street addresses
  • Web addresses and IP addresses
  • Social Security numbers and other labeled ID numbers
  • Ages 90 and over, hospital and clinic names

Keeps

  • The clinical content: history, exam, vitals, labs, meds, assessment and plan
  • Bare years ("quit smoking in 1998") and ages under 90
  • Employer and vendor names, flagged for review and kept unless you remove them
  • "The only..." descriptions, flagged for review and kept unless you remove them

Built around the 18 HIPAA identifiers

  • A clinical de-identification model (a BERT classifier trained on hospital reports) reads the note. Pattern rules cover what models miss: MRNs, NPI and DEA numbers, insurance IDs, web addresses. Together they go after the HIPAA identifiers that show up as text. Biometrics and photos never do, and the catch-all (any other unique number or code) only sometimes does.
  • On a public benchmark of 2,972 identifiers in clinical text it missed 66 (one surviving character counts as a miss) and touched 10 of 219 sentences that had none.
  • It still misses relative dates ("two days after surgery"), a facility name run together with a city, a name split across a very long note, scanned PDFs and photos, and notes in languages other than English.
  • Scrubs finds the identifiers and removes them. Calling the record de-identified is a legal call, and that stays with you.

How it runs

  1. 01First visit: a one-time download of about 92 MB. On a computer it starts as soon as the page opens; on a phone it waits until you add a note. After that the model stays in your browser.
  2. 02Scanning a note takes a few seconds on a laptop. A full H&P can take half a minute.
  3. 03Nothing you paste leaves your computer. The page makes no request that carries your text, and there is no analytics on it.
  4. 04Every flagged item can be kept or removed, and you can select text it missed and mark it yourself. Repeated values get the same placeholder, like [PATIENT_1].

CasePanel

When you need the whole chart read

Scrubs is for the bits you'd paste anyway: a plan you want sanity-checked, a paragraph you want tightened. When you want the entire chart read, prior admissions included, that's CasePanel. It works under a BAA on every plan, and Casey drafts the H&P, the admission orders and the med rec for you to review and sign.

See how CasePanel reads a chart

Common questions

Is ChatGPT HIPAA compliant?

No, for the plans most people use. The Free, Plus and Pro plans have no BAA, and OpenAI says content from those plans may be used to train its models unless you opt out. The plans with a BAA, as of October 2026: ChatGPT for Healthcare, ChatGPT Enterprise with Regulated Workspace, ChatGPT for Clinicians and the API with Modified Retention. A BAA covers OpenAI's side. Whether you can paste a patient's note is your hospital's call.

Is it fully local?

Yes. The model and the pattern rules run inside your browser tab, so your note never leaves the computer. The only downloads are the model and the page's own code, and there's no analytics on this page.

Is this de-identification?

No. It's a PHI scrubber: it finds the identifiers and removes them. Calling the record de-identified is a legal call, and that stays with you. Under the hood: a clinical de-identification model (a BERT classifier trained on hospital reports) plus pattern rules for MRNs, NPI and DEA numbers, insurance IDs and web addresses, covering the HIPAA identifiers that show up as text (biometrics and photos never do, and the catch-all only sometimes). On a public benchmark of 2,972 identifiers in clinical text it missed 66, counting one surviving character as a miss, and touched 10 of 219 sentences that had none. Read the cleaned text before you send it.

What counts as PHI?

Anything in the record that could point back to the person. HIPAA's list of 18 identifiers, the ones its Safe Harbor method strips, in chart terms:

  1. Names
  2. Address parts smaller than a state: street, city, county, ZIP
  3. Dates tied to the patient, except the year, and ages 90 and over
  4. Phone numbers
  5. Fax numbers
  6. Email addresses
  7. Social Security numbers
  8. Medical record numbers
  9. Health plan and member IDs
  10. Account numbers
  11. Certificate and license numbers
  12. Vehicle IDs and plates
  13. Device IDs and serial numbers
  14. Web addresses (URLs)
  15. IP addresses
  16. Biometrics (fingerprints, voiceprints)
  17. Full-face photos
  18. Any other unique number, characteristic or code

Scrubs goes after the ones that show up as text. Biometrics and photos aren't text. The catch-all sometimes is, and the pattern rules catch its common shapes.

Does it work on text that isn't a clinical note?

Yes. Names, phones, emails, addresses, SSNs and card numbers get caught in any text, so it works on PII too. The clinical model adds the chart-specific ones: MRNs, dates of service, facility names.

What does it miss?

Relative dates ("two days after surgery"), a facility name run together with a city, a name split across a very long note, scanned PDFs and photos, and anything not in English. It also can't tell you whether a story with no name in it still identifies someone. Read the cleaned text before you copy it.

Can I use it offline?

Yes, after the first visit. The app and the model stay in your browser cache. If the browser has cleared the model, Scrubs runs on patterns only until you're back online and it can download the model again.

What about dates and ages?

Dates go, bare years stay. Ages 90 and over go, ages under 90 stay. Any date with a month or day in it is removed; a year on its own ("quit smoking in 1998") is not on HIPAA's list and is kept. Relative dates in prose ("three weeks ago") are not caught.

Why is the first load 92 MB?

Because the model comes to your browser instead of your note going to a server. That is the model (70 MB) plus the software that runs it (22 MB), downloaded once and kept by your browser.

What about scanned PDFs?

They don't work. A scan is an image and there is no OCR. A PDF with selectable text is read on your machine like pasted text, and the export is cleaned plain text, not a redacted PDF.

Aclera Scrubs is a PHI scrubber. It removes identifiers from HIPAA's list of 18 in text you choose to share. It does not certify a record as de-identified under 45 CFR 164.514(b). You review the result and stay responsible for what you send and where. No text reaches Aclera: the model runs in your browser, and the page makes no request that carries your note. Built on open-source software: third-party licenses and notices.