跳转至

African context-sensitive attributes

OpenMed's deterministic safety sweep includes a small, extensible seed set for African healthcare text. It covers named healthcare facilities, mobile-money references, and explicitly contextualized ethnic or tribal affiliation mentions. The seed is intentionally not a pan-African gazetteer and contains only synthetic test data.

This feature is technical decision support, not legal advice. Organisations remain responsible for choosing a lawful basis, applying local requirements, and reviewing whether a broader or narrower redaction posture is appropriate.

Runtime mapping

Context shape Canonical label Policy class Default posture
Named clinic, mission hospital, dispensary, health centre, or referral facility ORGANIZATION QUASI_IDENTIFIER Non-keep in the six African profiles
M-Pesa, MTN MoMo, Airtel Money, and other configured wallet transaction or account references ACCOUNT_NUMBER DIRECT_IDENTIFIER Non-keep in the six African profiles
Race, ethnicity, or tribal-affiliation mention with explicit nearby context ETHNICITY SENSITIVE_ATTRIBUTE mask in the six African profiles and strict_no_leak

Terms, context words, regex templates, and profile defaults live in openmed/core/data/africa_context_terms.json. The merger only renders those data entries into its existing PIIPattern configuration. Ethnic-affiliation seed terms require nearby context such as ethnicity, tribe, or identifies as; a standalone term is not sufficient. Facility matching likewise requires a capitalized facility name plus a configured healthcare-facility cue. These constraints reduce unrelated matches while preserving the no-leak posture for planted fixtures.

Statutory basis

The mappings are conservative engineering controls informed by official legal texts and, where noted, a reference translation:

Extending the seed set

Add operators, facility cues, affiliation terms, or regional variants only in africa_context_terms.json. Keep additions synthetic, explain the regional scope, and add a planted fixture that proves both detection and redaction. Do not add real patient, facility-account, wallet, or transaction data. Avoid broad affiliation terms without contextual gating because many group names are also languages, places, surnames, or ordinary clinical-note content.