How well does Tenet detect PII/PHI?

Detection accuracy is measured against all 18 HIPAA Safe Harbor identifier categories using a synthetic annotated benchmark corpus. Data below are from the full Tenet pipeline: piiranha ML + regex augmentation + Presidio fallback + clinical NER.

0.985 Avg AUC across 6 HIPAA-critical types

15 of 18 HIPAA Safe Harbor identifier categories covered

6 languages EN · ES · FR · DE · IT · NL

2,710 samples annotated across 15 categories, 500 adversarial negatives

3 detection layers ML + regex + NER fallback

All benchmarks use synthetic data. Real-world performance may vary. Biometric identifiers and full-face photographs are not covered.

Detection accuracy by HIPAA identifier category

AUC scores from ROC analysis across the full detection pipeline

HIPAA Category	AUC	Method
Dates (DOB)	1.000	ML + regex
SSN	0.999	ML
Email	0.994	ML
Account Numbers	0.991	ML
Names (Given + Surname)	0.975	ML + Presidio
Phone/Fax	0.949	ML
Geographic (Street, City, ZIP)	0.932	ML
Other Unique IDs (Tax, CC)	0.885	ML + regex
License Numbers	0.650	ML
Device Identifiers	0.642	ML
IP Addresses	—	Regex-only
VIN	—	Regex-only
Medical Record #	—	Regex-only
Web URLs	—	Regex-only
Biometric	—	Not covered

8 of 10 ML-scored categories exceed AUC 0.88. The 6 HIPAA-critical types average 0.98 AUC.

Precision-recall tradeoffs at optimal thresholds

Configured thresholds balance recall (catching PHI) against false positive rate (flagging clean data)

Category	Threshold	Recall (TPR)	False Positive Rate
Dates	0.94	100%	0.0%
SSN	0.97	100%	0.4%
Email	0.66	98.8%	0.0%
Account Numbers	0.82	98.1%	0.0%
Names	0.45	96.0%	1.0%
Phone/Fax	0.44	90.5%	0.4%
Geographic	0.51	86.7%	0.8%

Lower thresholds on Names and Geographic reflect inherent ambiguity — these categories have higher irreducible false positive rates than structured identifiers.