PHI detection
How Sentinel detects protected health information in a prompt and redacts it before the prompt is sent
The SUPERWISE® Sentinel PHI guardrail detects protected health information (PHI) in a prompt — patient names, clinical details, and health identifiers — and redacts each value before the prompt goes to the model provider. The request still proceeds, and the third-party model never sees the health information.
Unlike PII redaction, which matches structured identifiers with patterns, PHI detection uses a machine-learning model that reads the surrounding context. That lets it catch health information that has no fixed format — a patient's name, a diagnosis, a medication, a hospital — which pattern matching misses.
Action: By default this guardrail remediates: each matched value is replaced with the guardrail's replacement text, REDACTED by default, which appears in the prompt as {{REDACTED}}. You can change the replacement text, or switch the guardrail to block matching requests instead. See Manage guardrails.
What does PHI detection cover?
Detection is grouped into categories you can turn on or off per Sentinel:
| Group | Examples |
|---|---|
| Personal identity & demographics | Patient and medical-professional names, date of birth, age, Social Security number, license, driver's license, and passport numbers |
| Location, contact & network data | Street address, city, state, country, ZIP code, phone, email, URL, IP address |
| Administrative, financial & health plan numbers | Medical record number, healthcare / national health number, insurance or policy number, medical codes, account, bank account, routing number, credit card |
| Physical, biometric & device serial numbers | Medical device IDs, serial numbers, vehicle identification numbers (VIN), biometric identifiers |
| Clinical & medical context | Conditions and diagnoses, injuries, blood type, drugs and dosages, procedures, and medical facilities |
Parameters
- Categories — choose which of the groups above to detect. Leave them all on to cover everything, or narrow the scan to the categories that matter for a Sentinel. See Manage guardrails.
- Confidence threshold — the lowest model confidence, from 0 to 1 (default
0.5), for a span to count as PHI. Raise it to redact only high-confidence matches; lower it to catch more.
Example interception
Inbound prompt from application:
"Summarize the chart for patient Jane Doe (DOB 1975-03-12, MRN 884-221), seen at Mercy General for hypertension and started on lisinopril 10 mg."What the upstream LLM actually sees:
"Summarize the chart for patient {{REDACTED}} (DOB {{REDACTED}}, MRN {{REDACTED}}), seen at {{REDACTED}} for {{REDACTED}} and started on {{REDACTED}} {{REDACTED}}."Because detection is model-based, matches vary with context and confidence rather than a fixed pattern. Use the confidence threshold to tune sensitivity, and the Test sandbox to see how your settings behave before you publish.
See also
- PII redaction: pattern-based detection of emails, phone numbers, and government and financial IDs.
- Secret & credential detection: keep API keys and tokens out of prompts.
- Manage guardrails: choose categories, set the action, and test before publishing.
Updated about 1 hour ago