La interfaz está localizada; las descripciones y documentos de las Skills se traducen automáticamente con IA a su idioma. El código, los comandos, rutas y términos técnicos se mantienen en el original, y prevalece el documento del autor.
Explorar Skills
Productividad · SKILL.md

deidentification-privacy

deidentification-privacy

Router for local-first PII and PHI extraction, de-identification, reversible re-identification, date shifting, surrogate generation, streaming/batch redaction, and audit-safe outputs.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentación de la Skill

GitHub

name: deidentification-privacy description: "Router for local-first PII and PHI extraction, de-identification, reversible re-identification, date shifting, surrogate generation, streaming/batch redaction, and audit-safe outputs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

De-identification and privacy

Use this sub-skill for local-first PII/PHI workflows that need:

  • extract_pii spans, labels, offsets, and confidences;
  • deidentify text redaction with masking, removal, replacement, hashing, date shifting, or format-preserving output;
  • reidentify when an authorized mapping is available;
  • policy profiles, surrogate generation, patient-keyed date shifts, audit-safe outputs, multilingual or code-mixed identifiers, and no-PHI logging;
  • document, streaming, and batch redaction over synthetic or authorized text.

Go here first

  • Need entity interpretation, model selection, or clinical grounding? Use sibling clinical-extraction-grounding.
  • Need table/row release risk or structured de-identification? Use sibling structured-risk-evaluation.
  • Need OCR or document intake before redaction? Use sibling multimodal-document-intake.
  • Need service, adapter, or FHIR/OMOP handoff? Use sibling interoperability-serving.

Core surfaces

  • Top-level APIs: extract_pii, deidentify, reidentify
  • Privacy helpers: openmed.core.anonymizer, openmed.core.date_shift, openmed.core.surrogate_vault
  • Long-text helpers: openmed.core.document_stream, openmed.core.streaming
  • Batch helpers: openmed.processing.BatchProcessor

Default operating rules

  • Keep inputs synthetic or already authorized.
  • Prefer local loaders, cached checkpoints, and explicit policy names.
  • Never log raw source text, mappings, prompts, or PHI-bearing outputs.
  • If the task needs re-identification, treat the mapping as sensitive as the source.
  • For long documents, stream or batch rather than loading everything into memory.

Bundled references

  • references/privacy-workflows.md
  • references/api-reference.md
  • references/troubleshooting.md
  • scripts/deidentify_synthetic_note.py
Skills guardadas