elisaterumi-ai/clinical-agent-skills/anonymization

When anonymizing clinical text, follow these steps:

1. Identify sensitive information

Detect and classify all Personally Identifiable Information (PII/PHI), including:

• Names (patients, doctors, relatives)
• Dates (birth, admission, discharge, events)
• Locations (cities, hospitals, addresses)
• Identification numbers (CPF, RG, SSN, medical record numbers)
• Contact information (phone numbers, emails)
• Organization names (clinics, hospitals, employers)
• Geographic details smaller than a state level
• Any unique identifiers that can re-identify a person

2. Apply anonymization

Replace each entity with a consistent placeholder:

• [PATIENT_1], [DOCTOR_1]
• [DATE_1]
• [LOCATION_1]
• [ID_1]
• [CONTACT_1]
• [ORG_1]

Rules:

• Maintain consistency: the same entity must always map to the same placeholder
• Do not invent or infer missing data
• Do not remove clinical meaning
• Use consistent indexed placeholders per entity type. If multiple entities of the same type exist, assign incremental IDs (e.g., [PATIENT_1], [PATIENT_2]).

3. Preserve clinical utility

• Keep all medical information intact:
• symptoms
• diagnoses
• medications
• procedures
• Maintain sentence structure and readability
• Avoid over-anonymization that removes useful context

4. Generalization (when needed)

If exact anonymization is not possible, generalize:

• Exact age → age range (e.g., "84 years" → "80+ years")
• Specific date → month/year or relative time ("March 2023" → "[DATE_1]")
• Precise location → broader region

5. Validate output

Before returning the result:

• Ensure no PII/PHI remains
• Ensure consistency of placeholders
• Ensure medical meaning is preserved

6. Output format

Return only the anonymized text.

Do not include explanations. Do not include metadata unless explicitly requested.