PHI redaction is the process of finding and removing Protected Health Information from text so less of it reaches downstream systems and people. Philterd builds open source, self-hosted PHI redaction software that runs inside your own network, so clinical notes, claims, intake forms, and patient messages are redacted before they reach a third party.
What counts as PHI
Under the HIPAA Safe Harbor method (45 CFR 164.514(b)(2)), data is de-identified only when all 18 identifiers are removed: names, geographic subdivisions smaller than a state, every date element tied to an individual, phone and fax numbers, email addresses, Social Security numbers, medical record numbers, health plan beneficiary numbers, account numbers, certificate and license numbers, vehicle and device identifiers, URLs and IP addresses, biometric identifiers, full-face photographs, and any other unique identifying number or code. A single clinical note can carry a dozen of these in a few sentences, which is why manual review does not scale.
How PHI redaction works
Effective PHI redaction combines two detection methods. Pattern-based rules (regex, checksums, and format validators) catch structured identifiers like MRNs, SSNs, and account numbers with high precision. NLP models catch the unstructured identifiers, including physician and patient names, hospital names, and locations, that no pattern can reliably match. Philter applies both, then redacts, masks, encrypts, or replaces each detection according to a policy you control. The practical trade-offs are covered in using an LLM or pattern-based rules for PII and PHI redaction.
Redaction is not the same as de-identification
Redacted text can still be PHI. Unless it meets a de-identification standard, HIPAA keeps applying to it, including the BAA you need with any vendor that handles it. Data de-identified by Safe Harbor or Expert Determination is no longer PHI, so the Privacy Rule no longer applies to it. Safe Harbor also requires that you have no actual knowledge the remaining information could identify someone, so automated redaction does not settle the question by itself. Most teams need both, applied to different workflows, and where the line falls is a decision for your privacy officer.
Self-hosted PHI redaction keeps data in your network
Where redaction runs decides who sees the text first. A hosted redaction API receives clinical text before it redacts anything, so that vendor needs a BAA. Philter and the Phileas library run inside your own VPC or on-premises, so what your policy detects is redacted before the text leaves that environment. The self-hosted software sends nothing to Philterd, so there is no BAA to sign with us for it. The detection logic is open source and can be inspected, and policies are versioned artifacts, so “what did we redact, and how, on this date” is answerable with a diff.
Where teams apply it
- EHR exports and research pipelines. De-identify clinical narrative before it lands in the analytics warehouse. Start from the HIPAA Safe Harbor policy.
- Clinical chatbots and RAG. Philter AI Proxy redacts the PHI it detects in prompts before they reach OpenAI, Anthropic, or Bedrock.
- Health-tech products. Embed redaction to give covered-entity customers an auditable answer on how you handle PHI.
See the full vertical picture in PHI redaction for healthcare, and how each capability maps to HIPAA and other regulations in the compliance matrix.
Get started
Philterd’s PHI redaction toolkit is free and open source under the Apache License, Version 2.0. Deploy Philter from your cloud marketplace into your own account, apply a HIPAA Safe Harbor policy, and measure precision and recall with Philter Scope before you wire it into production. For personal data beyond health records, see PII redaction.