De-identification Engines

DelPHI leads with delphi-om, our clinical-NER transformer. Other engines remain available as faster local fallbacks or for comparative analysis.

Quick comparison

Engine	Type	Accuracy	Speed	Privacy	Best for
Delphi-OM (v2) Default	Transformer NER	95.8% recall (benchmark)	Medium	Remote inference	Default choice for clinical documents — charts, notes, discharge summaries — where accuracy matters more than throughput.
Microsoft Presidio	Rule-based	Pattern coverage	Fast	Local	High-volume preprocessing or environments where remote inference is not allowed.
Hybrid Ensemble	Multi-engine	Highest (ensemble)	Slow	Local	Comparative QA and one-off audits where you want maximum recall regardless of cost.
MedSpaCy	Medical NER	Medium	Medium	Local	Clinical-section-aware processing on the local machine.

Delphi-OM (v2)
Transformer NER
Default

Three-stage pipeline. (1) The delphi-om clinical NER transformer running on a hosted inference endpoint proposes candidate PHI spans. (2) DelPHI's proprietary clinical-code shield post-filters those spans — preserving ICD-10-CM / HCC / CPT / NDC codes and a curated 151-term clinical vocabulary that a general NER would otherwise over-redact. (3) Redactor applies the chosen mode. The shield and redactor are DelPHI proprietary. This is the default engine.

Strengths

•Strongest clinical NER coverage of the available engines
•Detects all 17 HIPAA Safe Harbor identifier types + catch-all
•DelPHI clinical-code shield (proprietary) preserves ICD-10-CM, HCC, CPT, NDC codes and a curated 151-term clinical vocabulary — model alone would over-redact these
•Three redaction modes — anonymise / pseudonymise / delete
•No local GPU required — inference runs remotely

Limitations

•First call is slow if the model is cold-loading on a serverless endpoint
•Requires DELPHI_NER_ENDPOINT + DELPHI_NER_TOKEN to be set on the backend
•DEVICE-identifier recall is the weakest type (~32% on the synthetic benchmark)

Best use case

Default choice for clinical documents — charts, notes, discharge summaries — where accuracy matters more than throughput.

Technical details

95.8% recall (benchmark)

Accuracy

Medium

Speed

Pipeline

delphi-om NER → clinical shield (DelPHI) → redactor (DelPHI)

Shield

Regex on ICD-10-CM / HCC / CPT / NDC + 151-term clinical vocabulary (proprietary)

PHI types

17 concrete HIPAA types + UNIQUE_IDENTIFIER catch-all

Modes

anonymise · pseudonymise · delete

Inference

Hosted inference endpoint (HF or Vertex AI)

Benchmark

95.83% overall recall on 60 synthetic notes (14/17 types at 100%)

Recommendations

Default — clinical accuracy

Use Delphi-OM for charts, notes, and discharge summaries.

Delphi-OM (v2)

Fully local / fast

Use Presidio when remote inference is not allowed or for high-volume preprocessing.

Presidio

Comparative QA

Use the Hybrid ensemble for one-off audits comparing engines side-by-side.

Hybrid