Ask a hospital for its data and you rarely get a clean table. You get exports from the HIS (HBYS), DICOM headers from PACS, lab results in one schema, nursing notes in another, and a folder of spreadsheets nobody fully trusts. The raw material for clinical AI is almost always there. Turning it into something a model — or a human analyst — can actually use is the real project.
The work happens before the model
Teams underestimate how much of a data project is plumbing. Before any analysis, scattered sources have to be ingested, matched to the right patient, de-duplicated, and normalized to shared units and vocabularies. A creatinine value means nothing until you know its units, its reference range, and that it belongs to the same encounter as the imaging study next to it. We treat this as first-class engineering, not a preprocessing footnote.
- Ingest from HIS/HBYS, PACS and laboratory systems without hand exports.
- De-duplicate records that describe the same patient or event.
- Normalize units, codes and terminology to one analysis-ready schema.
- Keep provenance, so every value can be traced back to its source.
De-identification is a pipeline, not a checkbox
"Remove the names" is where careless projects stop and where real ones start. Identifiers hide in free-text notes, in accession numbers, in rare diagnoses that single a patient out, and burned into image pixels. A defensible pipeline separates direct identifiers, pseudonymizes what must be linkable, and suppresses or generalizes the quasi-identifiers that re-identification attacks rely on — and it does this the same way every time, with logs.
KVKK by design, not as an afterthought
Under KVKK, health data is a special category with a high bar for processing. That is easier to meet when privacy is built into the architecture rather than bolted on at the end. In practice that means data minimization — only the fields the purpose actually requires — clear purpose limitation, anonymization or pseudonymization as the default, and processing that can run on the institution's own infrastructure where the data must not leave the building.
Clean, structured, de-identified data is not the boring part of clinical AI. It is the part that decides whether the interesting part is allowed to exist.
Get the data foundation right and everything downstream — decision support, imaging, automation — gets faster and safer. Skip it, and you are building on sand that regulators, and clinicians, are right not to trust.