Human-subject data, identifiers, and what de-identified actually requires

1 · Learn the move · Fact-check list

"De-identified" is a defined status under your IRB protocol and the rules that govern your study - not a feeling you get when the name column is gone. Direct identifiers are the obvious layer. The layer that catches people is indirect: dates of service, rare diagnoses, small demographic cells, free-text fields where a nurse typed a name, and combinations - zip code plus age plus condition can be a fingerprint. The working move is a fact-check list: have the model build an identifier checklist from the categories your rules define, then screen a described dataset against it, field by field. The screen only ever produces flags and questions. It is never a clearance: the determination that a dataset is de-identified belongs to your IRB and your institution's process, and nothing in this course - and no model output - substitutes for it. And none of this is patient-care guidance; that is a different profession's job.

Build an identifier screening checklist for human-subject data: direct identifiers, indirect identifiers, dates, geographic detail, free-text risk, and small-cell combinations. Then screen this dataset description against it, field by field - I will describe fields only, never paste records: [field list with types]. Output: FIELD, RISK CLASS, FLAG or CLEAR-ON-DESCRIPTION, and a question list for my IRB. State plainly that a clear on this screen is not a de-identification determination.