mkocher/hipaa-phi-detector

HIPAA PHI Detector (DistilBERT)

A fine-tuned DistilBERT model for detecting Protected Health Information (PHI) in text, covering all 18 HIPAA Safe Harbor categories.

Model Details

• Architecture: DistilBERT (66M params) with token classification head
• Training: Fine-tuned on 5,000+ synthetic HIPAA examples
• Labels: 37 BIO labels (18 entity types x 2 + O)
• Framework: PyTorch / HuggingFace Transformers

Supported Entity Types

Label	HIPAA Category
NAME	Names
LOCATION	Geographic subdivisions
DATE	Dates
PHONE	Phone numbers
FAX	Fax numbers
EMAIL	Email addresses
SSN	Social Security numbers
MRN	Medical record numbers
HEALTH_PLAN	Health plan beneficiary numbers
ACCOUNT	Account numbers
LICENSE	Certificate/license numbers
VEHICLE	Vehicle identifiers
DEVICE	Device identifiers
URL	Web URLs
IP	IP addresses
BIOMETRIC	Biometric identifiers
PHOTO	Photographic images
OTHER	Any other unique identifying number

Usage

from transformers import pipeline

pipe = pipeline("token-classification", model="mkocher/hipaa-phi-detector", aggregation_strategy="simple")
results = pipe("Patient John Smith, SSN 123-45-6789")

Or with the aare-core package:

from aare import HIPAAGuardrail

guardrail = HIPAAGuardrail()
result = guardrail.check("Patient John Smith, SSN 123-45-6789")
if result.blocked:
    print(f"PHI detected: {result.violations}")

License

Apache 2.0