De-identifying a transcript: how to replace details and preserve meaning
A practical workflow for de-identifying transcripts: inventory identifiers, apply consistent replacements, check indirect clues, store the mapping separately and review the result.
Deleting a name is insufficient if other details can identify someone
Interviews may contain phone numbers, addresses and surnames, as well as a rare job title, the exact date of an event, a small team’s name or a distinctive story. Even without a name, those details can point to someone. Start by mapping identifiers and identifying who will receive the text.
Distinguish de-identification from pseudonymization. If a lookup table can restore the name, the text can still be linked to the person and requires protection. This editorial workflow is one part of the work. Assess the applicable law, contract and organizational policies separately.
Prepare a copy and check the risk of re-identification
Define the purpose and audience
Record why you need the text, who will see it and which details are necessary for analysis. The wider the audience, the more rigorous the review should be. Collect only the information you need.
Map the identifiers
Mark direct identifiers: names, contact details, document numbers and addresses. Then look for indirect clues: a rare role, exact dates, location, a one-off project or family relationships. Review individual fields and combinations.
Apply consistent replacements
Use the same label for each participant throughout the set, such as [Participant A] or [City]. Generalize only as far as the task allows. Work on a copy and store the original according to the applicable access and retention rules.
Have someone review the copy independently
Ask an authorized colleague to read the text as the intended recipient would. Check headings, filenames, comments and metadata as well as the visible text. Record the version, decisions about ambiguous details and who can access the copy afterward.
Interview with an employee at a small branch
This is a fictional training example. The names, circumstances and outcomes are included solely to explain the method.
I’m the only night-shift engineer at the branch on Severnaya Street. On 14 March, I handled an emergency with the director.
Removing the name does little here: the rare role, address and date can single out the employee. In a training copy for analysis, you might write, “[Technical employee] at [regional branch] described responding to an incident in spring.”
Keep the detail the study needs. If working the night shift matters, keep that and generalize other details further. This is a fictional example. The decision depends on the audience and the context of the dataset.
| Data type | Example | Action |
|---|---|---|
| Direct identifier | Name, phone number | Remove or replace |
| Indirect clue | Rare job title | Generalize or omit |
| Replacement key | A → named person | Store separately |
| Metadata | Filename, comment | Check before sharing |
De-identification record for a working copy
Complete the record without copying unnecessary personal data. Keep the replacement key in a separate, secured location.
Purpose: Recipients and access level: Source version: Working copy: Direct identifiers: Indirect clues: Replacements applied: Details kept and why: Where the key is stored (omit the key itself): Independent reviewer: Result date and version: Review date or trigger:
Check the text and its surrounding information
- The purpose and audience are defined and limited.
- Direct identifiers and indirect clues have been checked.
- The replacement key is stored separately from the copy.
- Headings, filenames and comments have also been reviewed.
A closer look
Is replacing surnames enough?
No. A role, date, location and distinctive event can identify someone in combination. Assess the details together.
Can I store the mapping table alongside the copy?
That increases the risk of reconnecting the text with a person. If you need the key, keep it separately with restricted access under your organization’s rules.