velo
ES EN PT
Request a POC

· 2 min read · NETBRAND CORP

The U.S. clinical identifiers generic NER misses: SSN, MRN, member IDs and more

Which identifiers in U.S. clinical documents generic named-entity models tend to miss, what their formats look like, and how to detect them reliably.

Why generic NER is not enough

A named-entity model learns what names and places look like in its training text. It has no concept of a medical record number or a member ID: those are just digits and letters. In a discharge summary they often appear right next to labels such as "MRN:" or "member ID", which is exactly the context a rule can use.

The identifiers to plan for

IdentifierTypical formatDetection approach
Social Security number412-58-7390Pattern with separators
Medical record numberMRN: 00482137Label trigger plus digits
Health plan member IDmember ID BRH-5582041Label trigger, alphanumeric
Prior authorization#88213-4"authorization" context
Provider NPINPI 1548392017Label plus 10 digits
Street address1452 Maple Ave, Apt 3BNumber, name, street type
ZIP codeVT 05401State code followed by 5 digits
Phone(802) 555-0142North American number patterns
Dates03/14/2026MM/DD/YYYY and written months

Provider identifiers such as the NPI are not patient identifiers under HIPAA Safe Harbor, but many teams redact them by policy because they narrow down where and when care happened. See Pseudonymization vs. anonymization under HIPAA, GDPR, LGPD and Colombia's Law 1581.

Context matters as much as format

  • Surnames repeat: "George Rogers (son)" shares a surname with the patient. A good system maps each real surname to one consistent fake surname, so the relationship survives synthesis.
  • Ages over 89 must be aggregated as 90 or older under Safe Harbor.
  • ZIP codes can keep their first three digits only where the area has more than 20,000 people.
  • Dates can be shifted by one offset per patient to keep intervals, but that is an Expert Determination technique, not Safe Harbor.

How to validate detection

Build a small annotated set from your own sources, then measure recall per identifier type, not a single global score. A model that finds 99% of names and 60% of member IDs is not ready. How to run a de-identification POC: 100 to 200 real documents and recall per entity describes the process.

Try it

The Velo portal runs a demo detector for U.S. clinical notes in your browser: paste a note with an SSN, MRN or member ID and watch it tokenize, synthesize or redact. The product itself combines clinical NER models with these rules, on-premise.

This article is general information, not legal advice.