On-premise PII redaction for clinical text: what it is and when you need it

What does "on-premise PII redaction" mean?

It is software that detects personally identifiable information (PII) and protected health information (PHI) in text, PDFs, images and audio, and then redacts, tokenizes or replaces it, all on hardware you control. Nothing is sent to a vendor's cloud to be analysed.

In a clinical setting the input is usually a discharge summary, a progress note, a referral letter or a scanned form. The output is the same document, or clean text, where names, identifiers, addresses and dates have been handled according to a policy.

How is it different from a cloud de-identification API?

A cloud API asks you to send the raw document to be cleaned. That can be acceptable with the right contracts, but it moves the most sensitive version of the data outside your perimeter, which is exactly what many hospitals, insurers and research groups want to avoid.

On-premise tools invert that: the raw document never leaves. Only the transformed output, if anything, goes out.

What does a good system detect in clinical text?

• Names of patients, relatives and clinicians, including repeated surnames in the narrative.
• National and health identifiers: in the U.S., SSN, MRN, health plan member IDs; see The U.S. clinical identifiers generic NER misses: SSN, MRN, member IDs and more for the full list per market.
• Addresses and locations, including ZIP codes and city names.
• Dates of admission, discharge and follow-up, and ages.
• Contact data: phone numbers and emails.
• Institutions: hospitals, clinics and insurers.

Generic named-entity recognition models catch names reasonably well, but they miss local identifier formats. Production systems combine a clinical NER model, rules for local identifiers and, optionally, a second pass with a local LLM.

When do you need it?

• You want to summarise or classify notes with a cloud LLM, but legal will not approve sending PHI. The pattern for this is in How to send clinical notes to a cloud LLM without sending the patient.
• You are building a dataset for analytics, training or research and need a documented de-identification step.
• You receive documents from many sources (native PDFs, scans, emails) and need one pipeline for all of them.
• Your privacy committee needs a measurable control, not a promise. See How to run a de-identification POC: 100 to 200 real documents and recall per entity for how to measure it.

What should you ask a vendor?

1. Does the engine run fully offline, including license validation?
2. Which identifiers does it support for my country, out of the box?
3. Can I choose per entity type whether to redact, tokenize, synthesize or shift dates?
4. Is the reversal map (the vault) stored only on my side?
5. Can I measure recall per entity type on my own documents?

Where Velo fits

Velo is an on-premise PII detection and redaction engine for clinical text in the United States, Colombia and Brazil. It runs on the Velo Nodo, a compact AI computer, or on your own Docker or Kubernetes cluster. You can watch a note being redacted live or request a POC.

This article is general information, not legal advice.