Protect patient privacy in every clinical note.
PHI identifies and de-identifies protected health information in unstructured clinical text through one automated pipeline - so your richest clinical data is safe to use across research, AI, analytics and partnerships.
Two coordinated stages: identify, then de-identify
PHI is built as two coordinated stages, deliberately separated so each can be scaled and governed independently - identification is the only GPU-bound component.
The result reads like a real note. Surrogates let PHI “hide in plain sight,” keeping the text natural and analyzable while the patient behind it stays protected.
Identify - a fine-tuned clinical-language model (~1.7B-parameter Qwen3 family, LoRA/QLoRA) tags every PHI element in narrative context, returning the note plus typed entities and character offsets.
De-identify - a second stage redacts each identifier or replaces it with a realistic, format-preserving surrogate - configurable per entity type and consistent within a document.
Patient John A. Smith, MRN 4471829, seen 03/14/2024 by Dr. Alvarez at Mercy General.
Patient Robert Lee, MRN 8830142, seen 06/22/2024 by Dr. Nguyen at Lakeside Health.
Hidden risk inside clinical text
Roughly 80% of clinical data lives in unstructured free text - the progress notes, radiology reports and narratives where the richest clinical signal sits. It is also where PHI is hardest to control: names, dates, MRNs, addresses and other identifiers are scattered through prose, easy to miss, and costly to remove by hand. Under HIPAA, a single missed identifier is a breach - and that risk keeps the most valuable clinical data sealed off from research, analytics and AI.
Why legacy de-identification tools fall short
PHI is buried in the narrative - identifiers woven into sentences, not tidy fields, so they are easy to overlook.
Manual review does not scale - human redaction is expensive, inconsistent, and can’t keep pace with millions of documents.
Legacy systems lack context - rule-based patterns and lookup tables miss nuance and mislabel identifiers.
Compliance exposure is binary - one leaked identifier is a breach, so organizations lock data down and block legitimate use.
A clean, composable de-identification pipeline
A clinical note flows through a clean, composable pipeline before emerging as safe, shareable text. The two stages are exposed as a simple contract: an entity-extraction call returns the note and its typed entities; an obfuscation call takes the note plus entities and returns the masked content.
Identify - the SLM tags typed entities with character offsets, in context.
De-identify - redact or replace each entity using a type-specific strategy.
Configurable - per entity type: realistic surrogate, generic redaction, typed redaction, or untouched.
Document-consistent - the same original value maps to the same surrogate every time, with casing preserved.
HIPAA Safe Harbor alignment
PHI detects 27 fine-grained entity types and obfuscates 30 in total. 16 of the 18 Safe Harbor identifiers are addressable in unstructured text; the two excluded - biometric identifiers and full-face photographs - are non-text by design.
Safe Harbor identifier coverage
| Identifier group | Representative entities | In scope |
|---|---|---|
| Names | Patient, provider, other names | Yes |
| Geography (sub-state) | Street, city/state, ZIP, country, facility | Yes |
| Dates & age over 89 | DOB, death, service dates, age 90+ | Yes |
| Telephone & fax | Phone numbers (fax handled as phone) | Yes |
| Email / Web & network | Email, URLs, IP addresses | Yes |
| Record / account numbers | MRN, account number, health-plan ID | Yes |
| Other identifiers | SSN, license, device, vehicle, ID/codes | Yes |
| Biometric identifiers | - (non-text) | No |
| Full-face photographs | - (non-text) | No |
Surrogate strategy by type
| PHI type | Surrogate strategy |
|---|---|
| Names | Random draw from curated name lists |
| Geography | Random draw from curated place / facility lists |
| ZIP codes | Format-preserving numeric substitution |
| Dates | Date shifting, optional truncate-to-year |
| Identifiers (MRN, account, SSN, device…) | Character/digit-class substitution preserving length & format |
| Phone / fax | Format-preserving number substitution |
| Email, URL, IP | Type-appropriate synthetic value |
Recall you can defend - measured the conservative way
Accuracy is measured as document-level All-or-Nothing recall - the share of documents in which every instance of an identifier type is caught - the conservative bar that matters for compliance. Benchmarked on radiology notes across eight imaging modalities, outperforming the Amazon Macie baseline.
Document-level recall (radiology-note benchmarks)
| PHI category | Document-level recall | Notes |
|---|---|---|
| Date of birth | Up to 100% | Highest on PET, PET/CT, ultrasound |
| MRN & patient names | ~98% | Consistent across modalities |
| Ages, dates & addresses | 96–97% | Mid-to-high 90s overall |
| Patient location (quasi-ID) | ~79% | Hardest category; active focus area |
Throughput & latency
| Stage | Throughput (min config) | p95 latency |
|---|---|---|
| PHI identification | ~100 requests / minute | ~10 seconds |
| Obfuscation | ~2,500 requests / minute | ~100 milliseconds |
Built for evidence, not generic claims
Accuracy, tuned to you - fine-tuned on your own notes and note types to reach 95%+ recall where it matters most.
Compliance you can prove - document-level scoring with statistical validation at 95% confidence, per note type - defensible evidence, not a generic claim.
Deploy your way - managed SaaS or fully on-premise; PHI never has to leave your environment.
Utility preserved - type-specific, format-preserving, document-consistent surrogates keep de-identified data usable for research and analytics.
Your notes never leave your control
PHI handles protected health information, so the architecture is built to keep that data locked down at every layer - from where the notes are processed to how the model itself is reached.
PHI stays in your environment - on-premise, notes are processed entirely inside your own infrastructure; nothing is sent out for inference.
The model is not directly exposed - clients interact only with the orchestrating API; the inference server is bound to localhost and isolated.
Hardened containers - distroless-style images limit shell access; the model server fails closed if it detects external exposure.
Encryption & mediated control plane - TLS in transit; an egress proxy mediates control-plane communication.
De-identify once - use the data everywhere
Once notes are reliably de-identified, the same dataset can safely serve many teams at once - health systems, researchers, payers, life sciences and health-tech vendors.
Research & real-world evidence
Share real patient data for studies, safely.
AI & model development
Train and validate models on real notes, compliantly.
Analytics & population health
Surface trends and outcomes without PHI exposure.
Data sharing & partnerships
Exchange data with vendors, registries and collaborators.
Non-production environments
Realistic data for testing, demos and QA with no live PHI.
Data products & secondary use
Build and license de-identified datasets.
De-identify once. Put your clinical data to work.
Book a demo and we’ll show PHI identifying and de-identifying on a sample of your own clinical notes.
Request a demo