Anonymize health data before handing it to AI

Health data is a “special category” under the GDPR: its processing is strictly framed, and pasting a clinical note into a consumer assistant means disclosing it to a third party. Before asking AI to rephrase a letter or summarize a file, replace the patient's identity, contact details, social-security number and identifying clinical elements with tokens. The model handles the text; no patient data leaves the browser.

Why health data isn't data like any other

The GDPR prohibits processing health data as a matter of principle, save for framed exceptions (care, explicit consent…). In the United States, the HIPAA “Safe Harbor” standard points the same way: it lists eighteen identifiers to remove for a record to count as de-identified. The shared message: before any external processing, strip what ties the information to a person.

• Direct identity: name, date of birth, address, social-security number.
• Care identifiers: file, insurance and admission numbers.
• Rare clinical elements that, alone, can re-identify (uncommon condition, precise dates).
• Contact details of the patient and their relatives.

What to remove before any prompt

1. 1The patient's identity and that of third parties (relatives, other patients named).
2. 2All administrative identifiers: social-security, file, insurance numbers.
3. 3Contact details: address, phone, email.
4. 4Precise dates and locations that, combined with context, re-identify.

A flow that preserves medical confidentiality

1. 1Detection: the engine spots identity, care identifiers and identifying elements.
2. 2Tokenization: each element becomes a neutral token, kept in local memory.
3. 3Sending: only the anonymized text goes to the AI — health data doesn't transit.
4. 4Restoration: the answer is de-tokenized in your browser, tied to the right file.

ONYRI Sanitize detects a file's identifying data — identity, social-security number, contact details, medical elements — and restores the answer in your browser. Care teams gain AI's help to rephrase or summarize, without ever exposing a patient or breaking medical confidentiality.