Guide · Core, Pathways and documents

De-identifying a case before you use an AI tool: what to remove and what to keep

Written for
Hospital doctors, residents and clinical teams
Reading time
6 minutes
Published
Updated

Written by the Sana Health team for clinicians. It summarises published guidance, is not a guideline and has not been peer reviewed. Check current local protocols before acting.

Key points

  • Most clinical questions can be answered without any information that identifies the patient.
  • Remove direct identifiers (names, record numbers, contact details, exact dates, addresses and photographs of faces) before a case goes into any external tool.
  • Indirect identifiers matter too: a rare diagnosis, an unusual job, a small town or a newsworthy event can identify someone in combination.
  • Keep the clinical facts: age, sex, a timeline in relative days, findings, results with units, medicines and doses.
  • Check attachments and images for headers, labels, barcodes, faces and burned-in text.
  • Your institution's policy and the law where you work decide what may be shared; when in doubt, ask your privacy or information governance office.

Why it matters

Patient-identifiable information cannot, in general, be shared outside the clinical care team without explicit consent [1]. Generative AI can reduce clinician burden, but it also brings risks that include privacy harms, misinformation and over-reliance on automation [2]. Removing identifiers before a case is entered into an external tool is the simplest way to reduce the privacy risk, and most clinical questions lose nothing in the process.

The legal frameworks differ by country and by purpose. The HIPAA Privacy Rule in the United States, the General Data Protection Regulation in the European Union and the United Kingdom, and the Privacy Act in Australia are examples, and each uses its own terms and definitions. Even expert recommendations on how to anonymise health data differ [3]. This article is a practical routine, not legal advice: your institution's policy, and the tools it has approved, take precedence.

What identifies a patient

Direct identifiers point to one person on their own. Indirect identifiers, sometimes called quasi-identifiers, do so in combination: a 34-year-old professional cyclist with a rare cardiomyopathy admitted after a race is identifiable without a name. Text can be de-identified, but the residual risk of re-identification from clinical free text is hard to quantify [4], and systematic approaches to reducing that risk are increasingly expected [5].

Remove or changeKeep or generalise to
Names, initials and nicknames, including those of relatives, carers and staffThe relationship only, such as her daughter or the GP
Hospital, insurance and national health record numbersNothing; they add no clinical value
Exact date of birthAge in years, or an age band for the very old, who are fewer and easier to identify
Calendar dates of admission, procedures and resultsRelative time, such as day 3 of admission or 2 weeks after surgery
Address, postcode, small town, workplace or schoolA region only when it matters clinically, for example for travel or endemic infection
Phone numbers, email addresses and social media namesNothing
A rare occupation, a public role or a newsworthy eventA general description, such as a manual worker exposed to solvents
Photographs of faces, tattoos and distinctive marksA cropped image of the relevant finding only

Keep the clinical facts

De-identification should remove identity, not meaning. Regulators who require anonymised documents have had to weigh the same trade-off between privacy and data utility [6]. For a clinical question the facts that matter almost always survive:

  • Age, sex and pregnancy status, and ethnicity only where it changes a reference range or a genetic risk
  • The presenting problem and its timeline, in relative days or hours
  • Observations and examination findings
  • Results with values, units and, where unusual, the laboratory's reference range
  • Imaging and other reports, with identifying headers removed
  • Medicines with doses, allergies with the reaction, and relevant background
  • Travel or exposure history at the level of country or region and relative dates

A routine that takes two minutes

  1. Write a fresh summary for the question you are asking rather than pasting from the record.
  2. Remove every direct identifier in the table above.
  3. Convert dates to relative time and the date of birth to an age.
  4. Generalise rare details unless they are the clinical point; if rarity is the point, remove everything else that narrows it down.
  5. Check attachments: letterheads, headers and footers, barcodes, file names, image metadata and names burned into images.
  6. Read it again as an outsider: could a colleague, a neighbour or a journalist recognise this patient?
  7. Record in the clinical notes, as local policy requires, that advice was sought and what was decided.

Manual de-identification of long narrative documents takes considerable effort, which is why automated methods have been developed for research datasets [7, 1]. For a single case, a careful manual routine applied to a short summary is usually enough, but no method removes all risk.

Cases that need extra care

  • Rare diseases and genetic findings, where the diagnosis itself narrows the population.
  • Children, particularly when safeguarding is involved.
  • Mental health, sexual health, substance use, HIV and termination of pregnancy, which carry particular stigma if disclosed.
  • Patients who are staff members, public figures or in custody.
  • Clinical photographs, in which ethical concerns about AI use include privacy and consent [8].
  • Small communities, where an occupation, a family structure or an event can be enough to identify someone.

What de-identification does not do

It reduces risk but does not eliminate it. It does not replace consent where consent is needed, or ethics and governance approval for research use of patient data. It does not make a tool appropriate for clinical use if the institution has not approved it. And it does not transfer responsibility: the clinician who asks the question remains responsible for what is shared and for every decision that follows.

If you are unsure whether a case can be shared, or which tools your organisation permits, ask your privacy, information governance or data protection officer before you start.

How Sana Health helps

The Diagnostic Reasoning engine asks for the facts that survive de-identification: age and sex, the presenting problem with its timeline, examination findings and observations, investigations with values and units, and medications, allergies and relevant background. None of these needs a name, date of birth, address or record number, and relative dates such as day 3 or two weeks after surgery serve its reasoning as well as calendar dates. Its report gives a ranked differential, the reasoning behind it, the red flags and the tests that would change the picture, followed by numbered references from PubMed.

Questions

Is removing the name enough to de-identify a case?

No. Dates, places, record numbers and rare combinations of details can identify a patient without a name. Remove the direct identifiers, convert dates to relative time, generalise rare details, and check attachments and images.

Can I paste a discharge letter into an AI tool if I delete the header?

Letters usually contain identifiers in the body as well as the header: dates, names of relatives and clinicians, addresses and hospital numbers. It is safer to write a short summary for the question you are asking, and to use only tools your institution has approved.

Which law applies to de-identifying patient data?

It depends on where you work and why the data are being used. The HIPAA Privacy Rule, the GDPR and Australia's Privacy Act are examples, and they define the terms differently. Your institution's policy and its privacy or information governance office are the authority.

References