This whitepaper proposes a practical framework for de-identifying clinical documents before AI-enabled medical record review in Colombia, where extraction, coding, audit and clinical search often need the full document. It is a non-empirical framework built from a narrative selection of public regulatory guidance and published literature, plus a fictitious clinical-document scenario; no patient records were collected or analyzed. Its taxonomy distinguishes 9 categories of sensitive content: direct identifiers, encounter identifiers, dates and times, institutions and locations, health professionals and relatives, rare clinical events, free-text social context, operational metadata and AI-generated output.
The proposed workflow keeps original processing copies, processed documents, mapping information and logs apart, with checks before AI review and after output generation, and a qualitative matrix organizes the controls into 10 areas for implementation planning. The authors stress that de-identification reduces identifying content but does not certify anonymity, and that Colombian law allows research use of sensitive data only with measures to suppress identity, without defining what suffices; no clinical anonymization standard was found among the sources examined. No re-identification risk, utility or detection performance was measured, so the framework needs evaluation under an authorized protocol. It is a whitepaper, not peer reviewed.
