Data quality in healthcare is whether the record in front of the clinician is accurate, complete, timely, consistent, and unique. It is not data governance (who owns the rules) and it is not interoperability (whether the pipe exists). Bad quality is how a correct interface still harms someone.
If you meant stewardship / MDM / catalogs → data governance in healthcare. If you meant moving the data to a new system → data migration best practices. If you meant the pipe → data integration in healthcare.
This page is the five dimensions and a framework you can actually run. Ignore vendor “2025 data quality report” padding.
Five dimensions
| Dimension | Means | When it fails |
|---|---|---|
| Accuracy | The value matches the real-world fact | Wrong allergy; wrong laterality on an image; billed procedure that was not done |
| Completeness | Required facts are present | No prior meds; missing outside labs; empty family history |
| Timeliness | Available when the decision is made | Critical potassium result posted next shift |
| Consistency | The same fact agrees across systems, including units | Weight in lb in the EHR and kg in the OR without conversion |
| Uniqueness | One person, one master record | Jon / John Smith split a medication list and an allergy |
They travel together. An accurate value in the wrong chart (uniqueness) is still a miss. A complete record that is twelve hours late (timeliness) is a different miss. Organizations that grade their own patient data as mixed or poor are describing this stack, not a single typo rate.
Where the errors come from
- Humans under time pressure — transposition, the wrong item in a long picklist, lb vs kg. Design and validation at entry beat a stern memo.
- Systems that do not agree — code systems, date formats, truncated fields on an HL7 feed. Interoperability can copy the error at scale.
- Process — no owner for the MPI, no audit of required fields, a migration that mapped “unknown” to a real code.
Clinicians who do not trust the chart stop using decision support and start calling the last hospital. That is how quality becomes a safety culture problem.
A framework that is not a slogan
- Name owners — stewards per domain (MPI, meds, labs, imaging). Governance writes the constitution; quality is the inspection.
- Validate at entry — required fields, range checks, allergy/med interaction prompts that clinicians do not click past because they fire on junk.
- Profile and audit — completeness, duplicate rate, code validity, lag from result to chart. Dashboards, not an annual project.
- Cleanse with a log — merge duplicates, standardise addresses, fix unit errors. Every transform is attributable.
- Close the loop — a one-click “this field is wrong” from the chart, and a visible fix. Staff who report into a void stop reporting.
Why AI does not forgive this
A model trained on mislabelled tumours, missing metadata, or a skin-tone slice that is not the population you serve will scale the error. Imaging is unforgiving: conflicting annotations and scanner-specific artifacts become the “ground truth.” Quality work on the training set is clinical risk management, not a data-science nicety.
Start where harm or money concentrates: allergies, meds, MPI duplicates, billing codes. A hospital-wide “fix data” programme without a first domain is how these projects stall.
PYCAD implements imaging pipelines on top of DICOM / FHIR / PACS — annotation and training included. It is not a data-quality platform for your EHR. Case studies.