Clinical note de-identification with local PHI redaction

What is clinical note de-identification?

Clinical note de-identification addresses information that can connect a medical narrative to a person. A name at the top of a letter is an obvious target. Less obvious details can appear in a family history, an appointment reference, a quoted message or a description of an unusual event.

Redaction is one technique used when preparing text: remove a value or substitute a label such as [email]. That operation alone does not establish that a note meets a legal de-identification standard. The remaining narrative, recipient and intended use also matter.

For a drafting task, begin with the smallest excerpt that can answer the question. Reformatting a paragraph may not require a full referral, historical notes or attachments. DataAnonymiser supports that preparation step; it does not decide whether a disclosure is authorised or whether an AI-generated clinical statement is correct.

What patient information should you review in a clinical note?

Check direct identifiers throughout the document: names, contact details, addresses, patient and insurance references, account numbers and appointment information. Look at headers, footers, copied correspondence and signatures as well as the main narrative. A reference repeated on page three can survive a review focused only on page one.

Then review relationships and context. A note may identify a relative or carer, name a workplace or describe a distinctive public event. Replacing the patient's name will not necessarily conceal who is being discussed when those details remain. Decide which relationships the task needs and which particulars can be omitted.

Dates, ages and chronology need deliberate handling. A date might be an identifying detail and also help explain the sequence of care. Do not casually replace it with an invented value. Apply the approved de-identification approach and have an appropriate reviewer check that the prepared text still serves its intended purpose.

This is a practical review guide, not an exhaustive legal checklist. Separate the question of what a detector can find from the question of what a recipient could infer. Clinical staff and information-governance reviewers bring context that a generic entity detector does not have.

HIPAA Safe Harbor and Expert Determination: where redaction fits

For organisations subject to HIPAA, HHS describes two routes to de-identification: Safe Harbor and Expert Determination. Safe Harbor requires removal of specified identifiers and no actual knowledge that the remaining information could identify the individual. Expert Determination requires a suitably qualified expert to assess a very small identification risk and document the analysis. [1]

HHS applies the identifier requirements to narrative free text as well as structured fields. Safe Harbor also has specific provisions for dates and ages over 89; deleting only a name and date of birth is not the full method. Consult the complete guidance when designing the process. [1]

DataAnonymiser is a redaction and review aid. It does not certify that an output meets Safe Harbor, perform an Expert Determination or make a workflow HIPAA compliant. Choose the applicable process with your organisation's privacy team before releasing a clinical record.

Patient-data redaction, pseudonymisation and GDPR

Under the European framework, pseudonymisation and anonymisation are distinct. The EDPB describes pseudonymisation as reducing the link to an individual, while anonymisation removes identifiability. Truly anonymous information falls outside EU data protection law. A note containing placeholders is not automatically anonymous. [2]

In a working session, a separate placeholder mapping can allow recognised tokens to be restored. That is useful for preparing and reviewing a draft, but retaining that connection matters when assessing the result. For the organisation able to reconnect the text to its patient, substituting labels should not be treated as permission to publish the record.

Do not use the US de-identification terminology as a shortcut to a GDPR conclusion, or vice versa. Have the responsible team assess the intended use under the applicable rules. A tool's processing location and a document's legal status are related questions, but they are not the same decision.

Example 1: restructure a referral without unnecessary contact details

The examples here use fictional details and illustrative substitutions. They demonstrate editing decisions, not measured detection results or records approved for disclosure.

Original excerpt: “Please contact Lena at lena@example.com to arrange the referral appointment. Her sister Maya will accompany her. An interpreter has been requested.”

Prepared excerpt: “Please contact [first_name_1] at [email] to arrange the referral appointment. Her sister [first_name_2] will accompany her. An interpreter has been requested.”

A narrowly framed request could ask an approved tool to organise the excerpt into contact arrangements and appointment requirements, preserving every placeholder and adding no new facts. Distinct labels help a reviewer follow who is attending without requiring the names during editing.

Before using the result, check whether the email placeholder or named relationship was needed at all. Review the original referral for other identifying context outside the excerpt. If values are restored later, verify that the patient and accompanying person remain correctly assigned; consistent tokens are not automatic identity resolution.

Example 2: shorten a discharge note without changing its meaning

Suppose the approved task is to shorten an administrative paragraph from a fictional discharge note: “Jonas requested written appointment instructions. His daughter Elena will collect the printed copy. Transport has not yet been arranged.”

Prepared text could replace the two names with [first_name_1] and [first_name_2]. The editing instruction should preserve the request, the collection arrangement and the unresolved transport status. It should not ask the model to infer that transport is booked or add advice about the patient's care.

Compare the proposed wording with the source before restoring names. “Transport has not yet been arranged” and “Transport has been arranged” differ by very little text and describe opposite states. Privacy review and accuracy review therefore need separate attention, even for apparently routine correspondence.

Where a clinical passage includes medications, doses, negation, uncertainty or event timing, a responsible clinician must review any transformation before clinical use. Redaction does not validate those details. If removing context makes the task unreliable, keep the work within the approved clinical environment rather than filling the gaps with invented information.

Example 3: prepare teaching material without recycling a recognisable case

Imagine a training coordinator who wants a generic exercise about arranging follow-up. A real note includes an unusual occupation, a specific event and a distinctive family relationship. Removing the names may still leave a story that colleagues or community members recognise.

For this task, a wholly invented scenario may be sufficient: “A patient requests written instructions and needs help arranging transport.” The exercise can ask learners to identify missing administrative information without reproducing the real person's history. The starting point is the teaching objective, not the maximum amount of source detail that can survive redaction.

Simply changing a real patient's name or blending details from several real records is not proof of anonymity. Review what the final narrative reveals and use the institution's approval process for case-based teaching. DataAnonymiser can help inspect and redact source text; it does not grant permission to reuse a patient's story.

A six-step workflow for redacting clinical notes before AI use

1. Define the task and approved destination. Specify whether you need formatting, administrative summarisation or another permitted use. Confirm the review responsibility and how the organisation authorises that use of patient information. Start with fictional text when evaluating a new workflow.

2. Select the minimum useful excerpt. Work from an authorised copy and keep the original clinical record intact. Exclude attachments, unrelated history and quoted correspondence unless they are necessary for the task. Record which source version the reviewer should compare against.

3. Run redaction and inspect the substitutions. Review the detected values, repeated identifiers and any residual-risk warnings. Where your plan supports custom terms, add specific references the detector missed. Do not assume a low-risk label means the output is legally anonymous or ready to release.

4. Review the remaining narrative and extraction quality. Look for recognisable circumstances and details about other people. Check that headings, dates, negation and units were read correctly. If a necessary detail cannot be retained within the approved disclosure conditions, change the task or its processing environment.

5. Submit only the approved version. Give the AI tool a bounded editing instruction and ask it to preserve placeholders and uncertainty. Check the selected account, attachments and pasted material before submission. Do not include the original note or its lookup mapping alongside the prepared text.

6. Compare, restore where needed and review again. While the relevant mapping is available, restore recognised placeholders and verify the recipient, identities and meaning. Treat the finished text as a new version requiring the normal review process, not as an automatically validated clinical record.

Scanned letters, PDFs and screenshots need an extra check

Document extraction introduces a step before redaction. A typed PDF, a scanned letter and a screenshot can yield different text even when they look similar on screen. Check the supported format and available extraction or OCR capabilities, then compare the extracted passage with the visible source.

Pay particular attention to repeated page headers, handwritten additions, tables and small print. A record number may be separated by a line break; two columns may appear in the wrong order. An extraction omission can hide a privacy issue or remove information needed to understand the passage.

DataAnonymiser's image workflow produces text for review. It does not blur the pixels of the source image, remove its metadata or create a visually redacted replacement scan. A screenshot or original PDF still needs its own handling decision. Share the reviewed text if that is the approved output, rather than attaching the original as supporting context.

Folder discovery is also a separate workflow from preparing a document. A DLP scan reports findings; it is not evidence that every file in the folder has been converted into a releasable clinical note.

Where patient data is processed, and where it goes next

In the Individual edition, DataAnonymiser processes content on the device. Enterprise uses a server the customer installs and operates within its own network, activated through explicit administrator configuration. Neither edition sends clinical content to DataAnonymiser-operated infrastructure for processing.

That boundary describes the redaction step. If you later paste the output into an external AI service, that is a separate disclosure decision. Review the actual service, account, contractual terms and permitted purpose. Local preparation does not make every destination appropriate for the remaining information.

For HIPAA-regulated workflows, HHS explains that a cloud provider handling electronic PHI on behalf of a covered entity or business associate is a business associate, requiring an appropriate agreement and compliance with applicable HIPAA obligations. Do not assume that applying placeholders removes those requirements when the information remains PHI. [3]

The desktop's working mapping is held in memory and is needed for restoration. Closing the app does not delete the original record, clipboard copies or documents saved elsewhere. Include those locations in the organisation's normal handling process, and keep patient-specific values out of reusable term presets unless their persistence is explicitly approved.

Evaluate accuracy and privacy with representative fictional notes

Before introducing the workflow, assemble fictional examples resembling the documents your team handles. Include repeated names, a patient and relative in the same paragraph, uncommon spellings, multilingual passages and scanned text. Record the intended substitutions so a reviewer can identify missed values and excessive redaction.

Assess usefulness as well as removal. Can the reviewer still distinguish the people, understand the sequence and tell what is confirmed, denied or uncertain? A document that hides identifiers but changes those relationships has not achieved the drafting task.

WHO warns that generative models used in health can produce inaccurate, biased or incomplete statements. This supports a separate check of the AI-produced result rather than treating successful redaction as validation of the subsequent answer. [4]

Start with one document type and one approved task. Agree who reviews difficult cases, how failures are reported and which conditions stop external sharing.