Clinical data types

Clinical data types for medical AI

MedCorpora can scope EHR extracts, radiology reports, laboratory results and multimodal clinical evidence alongside imaging when fields are lawfully available, technically linkable and relevant to the model task. Each data type requires its own privacy, terminology, missingness and provenance controls.

Published 30 July 2026 · Reviewed 12 August 2026 · MedCorpora
StructuredDiagnoses · labs · medications
TextReports · notes
ImagingDICOM · derivatives
LinkagePatient · encounter · time
01

Clinical context changes what an image means

Imaging alone may not establish diagnosis, severity, treatment or outcome. Linked clinical evidence can strengthen a dataset when it is temporally correct, sufficiently complete and covered by the approved programme.

02

Data types are not interchangeable

A diagnosis code, report impression, pathology result and longitudinal outcome represent different evidence. Dataset cards should identify the source and transformation of every label.

  • Structured EHR fields and encounter history
  • Radiology reports and impression text
  • Laboratory results with units and reference ranges
  • Pathology, procedures, medications and outcomes where appropriate
  • Imaging and clinical timelines joined under pseudonymous identifiers
03

Minimize before release

Only fields necessary for the defined programme should move into an external release. Direct identifiers and unnecessary free text remain outside the approved product.

04

Questions, answered directly.

Can imaging and EHR data be linked?

Potentially, using approved pseudonymous relationships and time windows when source systems support reliable linkage.

Is every clinical field complete?

No. Missingness and source-system coverage must be measured rather than assumed.

Can free text be released?

Only after appropriate authority, privacy processing and programme-specific review.

Institutional engagement

Define the cohort.