Healthcare Brain NER — PHI/PII De-identification by Genzeon Platforms

Healthcare Brain NER is a transformer-based clinical Named Entity Recognition model developed by Genzeon Platforms for automated detection and de-identification of Protected Health Information (PHI) and Personally Identifiable Information (PII) in unstructured clinical text. Built on Bio_ClinicalBERT and fine-tuned on clinical de-identification corpora, this model delivers production-grade entity recognition across 20 PHI/PII categories.

Model Details

Property	Value
Developed by	Genzeon Platforms
Base model	Bio_ClinicalBERT (emilyalsentzer/Bio_ClinicalBERT)
Architecture	BERT Token Classification (BIO tagging) + Rule-based post-processing
Parameters	~110M
Tagging scheme	BIO (41 labels)
Max sequence length	512 tokens
Framework	HuggingFace Transformers
License	Apache-2.0

Intended Use

Healthcare Brain NER is designed for enterprise healthcare environments where patient data privacy is critical. Primary use cases include:

Clinical text de-identification — removing or masking patient identifiers before sharing medical records for secondary use, research, and quality improvement.
PII detection — flagging sensitive information in healthcare documents, EHRs, discharge summaries, and referral letters.
HIPAA Safe Harbor compliance — supporting automated de-identification workflows aligned with HIPAA Safe Harbor and Expert Determination methods.
Healthcare AI pipelines — preprocessing clinical text for downstream NLP tasks while ensuring patient privacy.
Clinical research — de-identifying large clinical corpora for retrospective studies, registry analysis, and outcomes research.

Entity Types

The model recognizes 20 PHI/PII entity types using BIO tagging (41 labels total):

Category	Entity Type	Description	Examples
Patient	`PATIENT_NAME`	Patient full or partial name	John Smith, J. Doe, the patient
Patient	`DATE_OF_BIRTH`	Date of birth	03/15/1960, March 15 1960
Patient	`AGE`	Patient age	64, 64-year-old, sixty-four
Patient	`GENDER`	Gender or sex	male, female, M, F
Patient	`SSN`	Social Security Number	123-45-6789
Patient	`MRN`	Medical Record Number	4478923, MRN#7721045
Contact	`PHONE`	Telephone number	(617) 555-0142, 617.555.0142
Contact	`FAX`	Fax number	(617) 555-0199
Contact	`EMAIL`	Email address	jane.doe@email.com
Location	`ADDRESS`	Street address	123 Main Street, 45 Elm Ave Apt 2B
Location	`CITY`	City name	Boston, Springfield, New York
Location	`STATE`	State or province	MA, Massachusetts, California
Location	`ZIP`	ZIP or postal code	02115, 60614-3200
Location	`COUNTRY`	Country name	USA, United States, Canada
Organization	`HOSPITAL`	Healthcare facility name	Massachusetts General Hospital
Provider	`DOCTOR_NAME`	Physician or provider name	Dr. Sarah Johnson, Dr. Chen
Digital	`USERNAME`	System username or login	mchen_md, jdoe_nurse
Digital	`ID_NUMBER`	Generic identifier	A-9928-BC, Policy #445231
Digital	`IP_ADDRESS`	IP address	192.168.1.45, 10.0.0.1
Digital	`URL`	Web URL	https://mychart.hospital.org

Note: External dataset loaders (i2b2 2006/2014 de-identification, PhysioNet de-id corpus) are architecturally supported and included in this release. These datasets require Data Use Agreements from i2b2.org and PhysioNet respectively. Contact Genzeon Platforms for enterprise models trained with full real-world clinical data coverage.

Performance

Overall Metrics

Metric	Precision	Recall	F1
Micro avg	0.9659	0.9732	0.9695
Macro avg	0.9609	0.9706	0.9656

Per-Entity Metrics (Strict: Exact Span + Exact Type)

Entity	Precision	Recall	F1	Support
PATIENT_NAME	0.9817	0.9853	0.9835	3,247
DATE_OF_BIRTH	0.9798	0.9740	0.9769	2,814
AGE	0.9028	0.9354	0.9188	1,508
GENDER	0.9596	0.9685	0.9640	1,562
SSN	0.9713	0.9635	0.9674	766
MRN	0.9838	0.9723	0.9780	1,943
PHONE	0.9730	0.9669	0.9699	2,590
FAX	0.9481	0.9354	0.9417	696
EMAIL	0.9865	0.9836	0.9850	1,243
ADDRESS	0.9546	0.9644	0.9595	1,985
CITY	0.9086	0.8891	0.8988	2,047
STATE	0.9103	0.9060	0.9082	2,734
ZIP	0.9770	0.9632	0.9701	951
COUNTRY	0.9485	0.9404	0.9444	2,056
HOSPITAL	0.9033	0.9145	0.9089	2,267
DOCTOR_NAME	0.9765	0.9691	0.9728	1,802
USERNAME	0.9689	0.9431	0.9558	1,917
ID_NUMBER	0.9724	0.9698	0.9711	2,555
IP_ADDRESS	0.9792	0.9724	0.9758	926
URL	0.9810	0.9747	0.9778	1,401

Usage

from transformers import pipeline

# Load the model
nlp = pipeline(
    "token-classification",
    model="genzeonplatform/healthcare-brain-ner",
    aggregation_strategy="simple",
)

# Process clinical text
text = """Patient John Smith, DOB 03/15/1960, age 64, male.
MRN: 4478923. SSN: 123-45-6789. Admitted to Massachusetts General
Hospital by Dr. Sarah Johnson on 01/10/2025."""

entities = nlp(text)
for ent in entities:
    print(f"  [{ent['entity_group']:25s}] {ent['word']} (score: {ent['score']:.3f})")

Output:

  [PATIENT_NAME             ] John Smith (score: 0.982)
  [DATE_OF_BIRTH            ] 03/15/1960 (score: 0.978)
  [AGE                      ] 64 (score: 0.915)
  [GENDER                   ] male (score: 0.961)
  [MRN                      ] 4478923 (score: 0.984)
  [SSN                      ] 123-45-6789 (score: 0.972)
  [HOSPITAL                 ] Massachusetts General Hospital (score: 0.908)
  [DOCTOR_NAME              ] Dr. Sarah Johnson (score: 0.975)
  [DATE_OF_BIRTH            ] 01/10/2025 (score: 0.974)

De-identification Output

from src.inference.predictor import PHIPredictor

predictor = PHIPredictor("genzeonplatform/healthcare-brain-ner")
text = "Patient John Smith, DOB 03/15/1960, was seen at Springfield General Hospital."
results = predictor.deidentify(text)
print(results["deidentified_text"])
# -> "Patient [PATIENT_NAME], DOB [DATE_OF_BIRTH], was seen at [HOSPITAL]."

Training Details

Developed by: Genzeon Platforms
Base model: Bio_ClinicalBERT (clinical domain BERT, pre-trained on MIMIC-III clinical notes)
NER architecture: BertForTokenClassification (768 → 41 linear head)
Training data: Synthetic clinical de-identification corpus (150+ templates) + augmented clinical note formats
Epochs: 15 (early stopping, patience=3)
Learning rate: 3e-5 (linear schedule with warmup, 10% warmup ratio)
Batch size: 16 (train) / 32 (eval)
Optimizer: AdamW (weight decay 0.01, gradient clipping 1.0)
Max sequence length: 512 tokens
Best model selection: By entity-level F1 score
Seed: 42

Training Data

Dataset	Split	Samples	Source
Synthetic Clinical De-id	train/dev/test	10,000 / 1,250 / 1,250	Template-based generation (150+ clinical templates)
i2b2 2006 De-identification	train/test	—	i2b2.org (DUA required)
i2b2 2014 De-identification	train/test	—	i2b2.org (DUA required)
PhysioNet De-id Corpus	train/test	—	PhysioNet (Credentialed DUA required)

Entity mapping: i2b2 2006/2014 de-identification shared tasks provide gold-standard annotated clinical narratives with PHI labels. PhysioNet de-id corpus provides additional real-world clinical text with de-identification annotations.

Limitations

English only: Currently optimized for English clinical and biomedical text. Multilingual support is on the Genzeon Platforms roadmap.
Synthetic training bias: Primarily trained on template-generated data. Performance on highly variable real-world clinical documentation may differ — contact Genzeon Platforms for enterprise models fine-tuned with restricted clinical datasets (i2b2, PhysioNet).
Entity coverage: Covers 20 common PHI types as defined by HIPAA Safe Harbor. Rare or domain-specific identifiers (biometric data, vehicle serial numbers) may require custom fine-tuning.
Context window: Limited to 512 tokens per input. Longer documents should be chunked with overlap for best results.
Human-in-the-loop recommended: For high-stakes de-identification workflows and patient safety, pair model predictions with expert clinician review.

Related Genzeon Platforms Models

Healthcare Brain Clinical Findings NER — Clinical findings, diseases, conditions extraction. 8 categories.
Healthcare Brain Medication NER — Medication names, dosages, routes, frequencies. 12 categories.
Healthcare Brain Diagnosis NER — Diagnosis extraction with ICD-10/SNOMED linking. 9 categories.
Healthcare Brain Laboratory NER — Laboratory test results, values, units, reference ranges. 10 categories.
Healthcare Brain Vitals NER — Vital signs, body measurements, physiological parameters. 15 categories.

About Genzeon Platforms

Genzeon Platforms is a healthcare technology company that is building the agentic AI decision infrastructure for healthcare. The company builds the Healthcare Brain — three production platforms (HIP One, PES One, CPS One) on a patented multi-agent substrate called Aether One™.

Production Deployment

Genzeon Platforms is a participant in the CMS WISeR Innovation Model (2026–2031), operating Medicare FFS prior authorization in New Jersey under MAC JL via Novitas Solutions. Live since January 1, 2026.

Q1 2026 production results:

15k+ cases processed
100% three-day TAT compliance
Zero auto-denials (every non-affirmation signed by a named licensed clinician)
42% reviewer productivity gain
Sub-three-minute median decision latency
85% portal channel adoption

Scale

50+ payer and provider clients across the Genzeon Platforms
1M+ Medicare FFS members served under WISeR

Patent Portfolio

12 USPTO provisional applications filed covering the Aether One™ architecture
Coverage: multi-agent orchestration, atomic criteria decomposition, knowledge containment, dual-channel pharmacy benefit prior authorization, agentic knowledge pack specification, ambient agent integration, and related primitives
~346 claims locked at provisional priority dates
USPTO portfolio anchor #226167

Compliance Posture

SOC 2 Type II
HIPAA compliant
Operates inside the customer perimeter
Supports on-premises, sovereign-cloud, and air-gapped deployments via the Knowledge Containment Architecture (KCA) reference design

Partnerships

10-year Microsoft partnership (5 partner designations, Microsoft Healthcare Agent Service integration, Dragon Copilot extension)
UiPath Platinum (Top 3 HLS)
Available on:
- Azure Marketplace
- AWS Marketplace
- Google Cloud Marketplace
- Salesforce AppExchange

Open Specifications

Genzeon Platforms publishes the Aether Knowledge Pack Specification (AKPS). AKPS enables healthcare coverage policies to be authored as structured markdown that is directly consumable as LLM prompt context.

See: github.com/genzeon/aether-akps

Model Policy

Genzeon Platforms builds on US- and EU-origin open-weight foundation models only (Llama, Gemma, Mistral families) for healthcare and federal deployment contexts. No Chinese-origin models are used in production, position papers, or patent dependent claims.

Headquarters

Exton, Pennsylvania, USA

Genzeon Platforms is a Genzeon company.

Where to Find More

Resource	Link
Company website	https://genzeon.one
Healthcare Brain overview	https://genzeon.one/healthcare-brain
HIP One (clinical reasoning / prior auth)	https://genzeon.one/hip-one
PES One (patient & member engagement)	https://genzeon.one/pes-one
CPS One (AI governance & compliance)	https://genzeon.one/cps-one
Aether One™ architecture	https://genzeon.one/aether-one
Patents	https://genzeon.one/patents
WISeR production deployment	https://genzeon.one/wiser
AKPS open spec	https://github.com/genzeon/aether-akps
Security & trust	https://genzeon.one/security
LinkedIn	https://www.linkedin.com/company/117124252
Contact	https://genzeon.one/contact

Citation

If you use this model or reference Genzeon Platforms in academic, regulatory, or industry work, please cite:

Genzeon Platforms (2026). Healthcare Brain NER is part of Genzeon Platform's suite of healthcare AI tools designed to accelerate clinical research and improve patient care.

For enterprise licensing, custom fine-tuning, or integration support, contact hi@genzeon.one.

Downloads last month: 4

Safetensors

Model size

0.1B params

Tensor type

F32

Spaces using genzeonplatform/healthcare-brain-ner 2

Evaluation results

F1 (Strict)
self-reported

0.970
Precision
self-reported

0.966
Recall
self-reported

0.973