Case studies · Healthcare

A hospital network de-identifies clinical text before it trains a model

Patient identifiers replaced with codes before clinical text reaches a model or a training set, with expert determination and a record of each step.

TowerRedactionClinical text before training
Data science to the training set

Discharge note for Asha Proto, MRN 00-418-27, admitted 03/02/2026, lives at 12 Elm St, Springfield. Stable on discharge.

Ready to send
Illustrative product view. Sample data, not client data.

The problem

Data scientists want clinical notes to train a model that drafts discharge summaries. Notes are full of names, dates, record numbers and addresses, and the privacy office will not release them as they are.

  • Beforethe model or training set sees the text
  • Codesin place of names, dates and record numbers
  • Recordedeach document and each replacement

Before Raidu

  • Notes held back from data science
  • Manual redaction by hand
  • No record of what was removed

With Raidu

  • Identifiers replaced with codes in their cloud
  • An expert reviews re-identification risk
  • Each replacement on the record

The challenge

The hospital’s data science team wants to train a model on clinical notes. Under HIPAA, notes can be used outside treatment and operations once they are de-identified: either every one of the 18 Safe Harbor identifiers is removed, or a qualified expert determines that the risk of re-identification is very small. Free-text notes rarely meet Safe Harbor on their own, so the privacy office chooses expert determination.

How they set it up

Raidu runs in The hospital’s own cloud. Notes pass through its redaction step, which replaces the identifiers the hospital’s policy covers, such as names, record numbers, dates and contact details, with codes before the text is written to the training set. A qualified expert reviews the method and a sample before release. Each document processed and each replacement goes on the record, so the privacy office can show what was done.

What this does and does not do

Redaction is one control inside the hospital’s de-identification process, not a substitute for it. The expert’s determination, the hospital’s policy and a business associate agreement with Raidu are what make the process hold.

Rollout

  1. Step 1Privacy office defines the identifiers to remove, starting from HIPAA's 18
  2. Step 2Raidu redacts the notes inside the hospital's cloud
  3. Step 3A qualified expert reviews a sample and the method (expert determination)
  4. Step 4The de-identified set is released for training

What they deployed

  • Redaction: identifiers the hospital's policy covers replaced with codes before text leaves their cloud
  • Evidence and Audit: a record of each document processed and what was replaced
  • A HIPAA de-identification process run by the hospital, with expert determination

Talk to us about your data and AI.

Tell us what you need to prove and to whom. We will show you the record it produces.

Talk to us