PHI de-identification for healthcare data lakes

Protecto finds PHI in claims tables and clinical notes and replaces it with tokens. Each patient keeps one token everywhere, so research, BI and AI keep working.

SOC 2 Type II · ISO 27001 · HIPAA, BAA available

Trusted by regulated enterprises & agentic platforms
Inovalon
Automation Anywhere
Citrix
Bank Of Muscat Logo
Dell
Case study result

How an insurance provider protects claims data and clinical notes

~750M

Claims documents in scope at a healthcare data platform

1,000+

Documents tested in under a week at Inovalon

50M+

Records handled at a healthcare insurer

Why PHI de-identification breaks healthcare analytics

A data team copies five years of discharge summaries into the lake to train a readmission model. A script masks the patient name column. The notes field still holds names, record numbers and a daughter’s phone number. Redaction then turns every person into the same placeholder, and the model can no longer tell who came back.

The copy is not safe, and it is not useful either.

PHI hides in free text.

Column rules miss the names, record numbers and phone numbers written inside notes, transcripts and logs.

Redaction cuts the patient thread.

When every identifier becomes the same placeholder, readmission counts, cohorts and links to claims stop working.

Teams go back to the raw extract.

Researchers and reviewers sometimes need the real patient. With no governed way to get it, raw copies spread and nobody can show who saw what.

One discharge note, three copies in your lake

Switch between the copies and see which questions your data team can still answer.

clinical_notes, one row

Discharge summary

What the data team can still do

    PHI values in the lake copy
    Same patient in claims
    Real values back

    Synthetic note. Your policy decides which entities are replaced and whether approved roles can reverse them.

    See this on notes of your own.

    HIPAA Safe Harbor identifiers, found in tables and notes

    Safe Harbor lists 18 identifiers to remove before health data counts as de-identified. Protecto detects them in structured columns and inside free text, and adds the IDs your organization uses, such as member numbers.

    You pick the outcome per dataset: tokens that approved roles can reverse, or anonymization with no way back. Your privacy team confirms each release.

    The 18 Safe Harbor identifiers

    1. Names
    2. Geographic units smaller than a state
    3. Dates, except the year
    4. Phone numbers
    5. Fax numbers
    6. Email addresses
    7. Social Security numbers
    8. Medical record numbers
    9. Health plan beneficiary numbers
    10. Account numbers
    11. Certificate and license numbers
    12. Vehicle identifiers
    13. Device identifiers
    14. Web URLs
    15. IP addresses
    16. Biometric identifiers
    17. Full-face photos
    18. Any other unique identifying number or code

    Source: HIPAA, 45 CFR 164.514(b)(2). Protecto works on text and imaging metadata. Photos and biometric data need their own controls.

    Pseudonymization or anonymization, dataset by dataset

    One policy engine, two outcomes. Choose by who will read the data and whether anyone should ever get the real values back.

    When the dataset is forUseWhy
    Analytics and model training inside your organizationPseudonymizationConsistent tokensEach patient stays distinct across tables, and nobody needs a real name to build a cohort or train a model.
    Care, claims or support teams who need the real patientPseudonymizationUnmask by roleApproved roles request the real value through a policy. The stored data does not change, and each request is logged.
    An LLM, RAG or search system reading clinical notesPseudonymizationTokens in free textThe model reads the clinical meaning, and patients, relatives and clinicians stay distinct.
    A dataset leaving your organization, such as a research partnerAnonymizationNo way backAn anonymization policy turns unmasking off for good, with no change to your pipeline calls.

    Tokenized data is usually still PHI under HIPAA. Your privacy team and counsel decide how each dataset is released.

    De-identify PHI in the pipelines you already run

    Call Protecto from the step that already moves the data, using its API, a Snowflake UDF, Databricks, Kafka or Spark. Tables keep their columns, so downstream jobs read them as before.

    For years of history already in the lake, submit an async job and collect the result when it finishes.

    What your team gets around the tokens

    Protecto finds the PHI, handles the volume, applies your policy and keeps the records your auditors ask for.

    01

    Detection in tables and free text

    50+ entity types across PII, PHI and PCI, plus custom entities such as member or encounter IDs, added in days.

    02

    Room for backfills and nightly loads

    Async APIs, queuing and batches handle years of notes and claims as well as each new load.

    03

    Real values only for approved roles

    Unmasking never happens on its own. It takes an explicit request, a policy that allows it and the right permissions.

    04

    Records for your auditors

    Event logs record each mask and unmask. Reports show where PHI flows and how your data policies were enforced.

    “Protecto enabled our team to harness the full power of AI and analytics while maintaining strict HIPAA compliance and data privacy.”

    AK
    Arun Krishnaswami
    Inovalon
    Security & compliance

    Built for regulated data at enterprise scale.

    Protecto combines data protection, audit logs, and ready-to-use policies for regulated data, whether deployed as SaaS, on-prem, in a private cloud, or air-gapped.

    SOC 2 Type II

    Audited controls

    ISO 27001

    Certified ISMS

    HIPAA

    BAA available

    GDPR

    EU data ready

    DPDP

    India privacy law

    CCPA / CPRA

    US privacy
    Independently verified to outperform AWS Comprehend and Microsoft Presidio · audit logs on every scan, mask, and unmask.

    Questions about PHI de-identification

    What is PHI de-identification?

    It removes or replaces the identifiers in health data, such as names, record numbers, dates and phone numbers, so the data no longer points to a person. Protecto does this in tables and in free text such as clinical notes, before the data is used for analytics or AI.

    Yes. Protecto detects the Safe Harbor identifiers in structured columns and free text, and you can add custom identifiers. Your privacy team still confirms each release. If you use Expert Determination, your expert reviews the output.

    Usually, yes. Tokenization is pseudonymization: while a mapping exists, a token can be linked back to a person. Many teams keep their BAA and access controls in place for tokenized data. Your counsel decides what applies to each dataset.

    Yes. Protecto finds PHI inside discharge summaries, progress notes, transcripts and logs, and replaces only the identifiers. The clinical words stay as written.

    Not with consistent tokens. The same value always gets the same token, so joins, cohorts and readmission counts still work, and patients stay distinct in free text. We measure how well AI answers hold up after masking with our accuracy metric, RARI.

    Run a backfill. Submit an async job, receive a tracking ID, check its status and read the result when it finishes. Async APIs need a paid plan. Book a demo and we will set up a proof of concept for your volume.

    Protecto runs on your servers or in your private cloud, or as SaaS. When Protecto hosts or manages the service, or has access to your installed instance, Protecto is a business associate and signs a BAA before PHI flows. When it runs in your environment with no Protecto access, Protecto does not receive PHI.

    Protecto runs in data pipelines through its API, a Snowflake UDF, Databricks, Kafka or Spark. The output is still data your tables can hold, so BI and ML tools read it as they do today.

    Yes. Protecto supports multiple tenants, each with its own policies, so one hospital, client or project never shares tokens or access with another.

    De-identify a sample of your own PHI

    Bring a few notes and a claims table, with synthetic data if you prefer. We will show you what Protecto finds and which queries keep working.

    Download Privacy Vault Datasheet

    This datasheet outlines features that safeguard your data and enable accurate, secure Gen AI applications.

    By submitting you agree to Protecto’s privacy policy.

    Cover of the Protecto Privacy Vault datasheet
    Datasheet

    Protecto Vault: sensitive data protection for AI workflows

    Features that safeguard your data across agents, AI applications, RAG pipelines, MCPs, data lakes, and enterprise systems.

    By submitting you agree to Protecto’s privacy policy.

    New: run any AI model without exposing real values.
    See how it works