De-identify clinical notes with human review

De-identify clinical text inside your own environment, then review every detection before anything is released. This walkthrough builds the cohort with the Assistant, scopes a document dataset, and runs the privacy review end to end.

Clinical Notes De-Identification with Human Review

Who this is for

Anyone who has to move clinical text somewhere it cannot carry PHI: research data offices, honest brokers and data stewards; privacy officers and compliance teams approving a release; AI and ML teams assembling training or evaluation corpora; analytics teams moving notes into a non-clinical enclave; and study teams sharing notes with an academic or commercial collaborator.

Why it matters

A research office agrees to send 5,000 clinical notes to an external biostatistics collaborator. HIPAA requires Safe Harbor or Expert Determination before those notes leave. Pediatric notes make this harder than usual: dates of birth carry age precision that matters, family member names appear throughout social history, and rare diagnoses can identify a patient on their own.

Pattern matching finds the medical record numbers. It does not find "Dr. Chen mentioned this to the patient's mother, Maria, at the Tuesday visit."

What you gain

De-identification that runs inside your environment, so the notes never leave your control, with a review stage before release. You see every detection, approve it, and correct what the model got wrong. That review is the artifact your privacy officer needs when they sign off, and it is the difference between a de-identification claim and a documented de-identification process.

Before you start

• Patients and clinical documents ingested
• A cohort defining the population whose documents you are releasing
• Access to Data Enrichment > De-Identification, and a named reviewer who will approve detections
• Your target standard settled in advance: HIPAA Safe Harbor or Expert Determination. The two lead to different review decisions

Step 1: Use the Assistant to create the cohort whose documents you will release

1. In the left navigation, go to Agents and Tools > Assistant.
2. Ask a cohort-building question based on the population whose documents you are releasing, for example: "Find patients with oncology notes from the last year."
3. Review the Assistant response, including the interpreted criteria and returned patient population.
4. If the patient population looks appropriate, click Create Cohort to begin cohort creation.
5. In the Advanced Filtering panel on the right side of the screen, apply additional filters if needed, such as diagnosis, document type, date range, provider, visit type, or demographic criteria.
6. Click Next to continue to Patient Management.
7. Review the patient list and exclude any patients that should not be included in the release cohort.
8. Click Next to continue to Cohort Details.
9. Enter the cohort name, description, category, and any linked cohort information.
10. Use a clear name that identifies the cohort purpose, such as external review, collaborator release, or privacy review.
11. Click Next to continue to Review Cohort.
12. Review the cohort configuration and optionally define a sample size.
13. Click Create Cohort to save the cohort.

Step 2: Build a document dataset from the cohort

1. In the left navigation, go to Data Curation Studio > Dataset Builder.
2. Select Cohort & Clinical Filter as the dataset mode.
3. Select Unstructured as the data type.
4. Click Next to continue to Cohort Selection.
5. Select the cohort created in the previous step.
6. Click Next to continue to Unstructured Criteria.
7. Apply document filters to limit the dataset to only the clinical documents that require privacy review.
8. Use filters such as note type, document date, source, provider, or visit context where needed.
9. Keep the dataset scope narrow enough for review, but broad enough to include all documents intended for de-identification.
10. Click Next to continue to Dataset Details.
11. Enter the dataset name and description.
12. Use a name that clearly indicates the dataset is intended for document de-identification or privacy review.
13. Click Next to continue to Review Dataset.
14. Review the selected cohort, document filters, and expected dataset contents.
15. Click Create Dataset to create the dataset.

Step 3: Confirm the documents that need privacy review

1. Open the created dataset from Dataset Explorer if a final check is needed before de-identification.
2. Review the included patients, documents, note types, dates, and source files.
3. Confirm that the dataset includes the clinical documents intended for privacy review.
4. Remove or exclude any documents that should not be part of the de-identification workflow, where the deployment supports dataset adjustment.
5. Confirm that the remaining dataset content matches the approved sharing scope.

Step 4: Run De-Identification on the document dataset

1. In the left navigation, go to Data Enrichment > De-Identification.
2. Start a new de-identification job or session.
3. Select the document dataset created for privacy review.
4. Select the appropriate de-identification profile and pipeline for text documents.
5. Review the job configuration before starting.
6. Click the start action to begin de-identification.
7. Monitor the job until processing is complete.

Step 5: Review detected sensitive information

1. Open the completed de-identification job details page.
2. Review detected sensitive information at the document level.
3. Inspect the original and de-identified text where comparison is available.
4. Confirm that patient names, identifiers, dates, contact details, locations, provider details, and other protected or sensitive elements are handled according to the review policy.
5. Approve, correct, mask, unmask, or flag detections according to the available review workflow.
6. Pay special attention to headers, footers, embedded tables, copied report sections, and repeated identifiers because sensitive information may appear outside the main clinical narrative.
7. Save review decisions where supported.

Step 6: Download reviewed de-identified text outputs

1. Confirm that the document de-identification review is complete.
2. Open the detailed de-identification result for the reviewed text document.
3. Download the reviewed output directly in JSON format.
4. Repeat the download process for additional reviewed documents as needed.
5. Store or share the downloaded outputs according to the approved governance, privacy, and customer-delivery process.
6. Document any known limitations, unresolved detections, or documents excluded from the review.

Recipe reference

Each stage of this scenario is also a reusable building block.

These steps have their own pages:

• Create a patient cohort with the Assistant
• Build a dataset from a cohort or ingestion results

De-identify a dataset before sharing or review

When to use on its own: Before sharing data externally, assembling a training or evaluation corpus, or reviewing images and documents that may contain patient identifiers.
Features involved: Dataset Builder, De-Identification, patient/document/image review, side-by-side original and de-identified review where available.
Edge cases / limitations: De-Identification requires a dataset; it isn't cohort-only. Export/save behavior should be confirmed before promising a complete handoff. Validate outputs before external sharing.
Value: Produces a documented privacy review a privacy officer can sign against before anything is released.