Anonymizing Health Data and Patient Texts for AI

Why health texts need special care

Health data belongs to the special categories of personal data under Article 9 of the GDPR, and medical staff are bound by professional secrecy. In the US, HIPAA governs protected health information. At the same time, clinical free text is where much of the useful information lives: discharge letters, radiology findings, nursing notes, referral letters. Hospitals, practices and research groups want to summarize, classify and analyze these texts with AI. The safest route is to remove personal information before the text reaches any model that runs outside your control.

This page explains general principles. It is not legal advice; involve your data protection officer and, for research, your ethics committee.

Common use cases

• Summarizing long patient histories for internal case conferences.
• Classifying free-text reports for quality management or incident reporting systems.
• Preparing text corpora for research projects or for evaluating AI tools.
• Teaching material and case studies for medical education.
• Drafting plain-language explanations of findings, with clinician review.

What has to go, and what is easy to miss

For orientation, the HIPAA Safe Harbor method lists 18 identifier types to remove, including names, geographic details smaller than a state, all date elements except the year, phone numbers, email addresses, record and account numbers, and device identifiers. HIPAA also allows an alternative route via expert determination. Under the GDPR, there is no checklist: data is anonymous only if a person can no longer be identified by any means reasonably likely to be used.

Free text adds challenges structured data does not have:

1. Embedded identifiers: "Patient was transferred from St. Mary's on the 14th after a fall at her bakery" contains a place, a date and an occupation.
2. Rare conditions: a rare disease in a small region can identify one person on its own.
3. Relatives and staff: names of family members, treating physicians and nurses appear throughout letters.
4. Dates and ages: exact admission dates and very high ages are identifying; shifting or generalizing them can preserve clinical meaning.
5. Headers and footers: letterheads, barcodes and record numbers in templates are often missed when only the body is processed.

Why local processing matters

If the anonymization tool itself is a cloud service, the original patient text leaves your infrastructure before it is cleaned. A tool that runs locally on a workstation or a hospital server processes the raw text without transmitting it. Only the cleaned text is then used with an external AI model, if at all. That also simplifies conversations with IT security and data protection.

The limits of anonymization

Be honest about residual risk. Automated tools miss things, and context can re-identify people, especially when the anonymized text is combined with other data. Pseudonymization, where a key allows re-linking, is useful for research but the data remains personal data under the GDPR. Practical safeguards:

• Use the minimum text needed for the task.
• Review samples, and review fully when texts are published or shared outside the organization.
• Restrict access to the anonymized corpus and document the process.
• Re-evaluate the risk when data is combined with other sources.

How jocapps can help

Textwash Pro removes personal data from text with a local AI that runs fully offline on your device, or works as a privacy layer in front of ChatGPT, Claude and co. It was built with researchers from UCL. If anonymization is one step in a larger process, for example classifying reports and routing them via SharePoint, kalAIDoCreate combines a hardened AI orchestration platform with senior AI architects to build governed workflows. Background: how to anonymize data before sending it to ChatGPT or Claude.

To discuss your use case, book a free 30-minute intro call.