Glossary · EU medical devices: MDR
Representativeness of training data (medical AI)
Also known as: Training data representativeness, Data representativeness, Representative training data
Representativeness of training data is the degree to which the data used to train, validate and test an AI model reflect the population, settings, devices and conditions in which the model will be used. Article 10(3) of the EU AI Act requires training, validation and testing data sets of high-risk AI systems to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose.
- EU
- USA
- AI
- Data integrity
In one sentence
Training data representativeness: how well an AI model's data reflect the patients, sites and devices it will meet; AI Act Art. 10(3) requires it.
Example
A sepsis prediction model trained on data from two university hospitals is checked against the intended use in regional hospitals: age distribution, laboratory devices and documentation habits differ, so the manufacturer adds test data from such sites and states the limitation in the IFU.
Explained in context
This term is a station of a mini course
- The medical-device evidence chain: from intended purpose to post-market evidenceStation: Training data provenanceStart the course
Context cards connect this term with others to answer one question. Also in British English and German.
- The medical-device evidence chain: from intended purpose to post-market evidenceHow do the parts of a medical device’s technical documentation connect — and what has to be in place before an AI function may answer from it?UKDeutsch
- Training data provenance is part of the device file, not a data-science recordWhy does the origin of training data belong in the technical documentation of an AI-enabled medical device?UKDeutsch
- Documenting the limits of a model: what the IFU of an AI-enabled device must sayHow should the limitations of an AI model be documented for the users of a medical device?UKDeutsch
How it applies
- Measured against the intended purpose: Representativeness is not a property of a data set alone. It is judged against the intended purpose: patient population, clinical setting, acquisition devices, geography.
- Timing in the EU: For AI in devices that need a notified body, the AI Act data governance requirements apply from August 2, 2028; the MDR already asks for clinical evidence that covers the intended population.
- USA: FDA's draft guidance on AI-enabled device software functions asks for a description of data sources, demographics and subgroup performance; it is a draft.
- Technical documentation: Record where the data came from (data lineage), which groups and settings it covers, and which it doesn't. State the uncovered groups as limitations in the IFU and the interface, not only in the data report.
Representativeness vs. data quality
A data set can be accurate, complete and well labeled and still unrepresentative — for example, clean data from one device model only. Data quality asks whether the data are right; representativeness asks whether they are the right data for the intended use.
External references
- Regulation (EU) 2024/1689 (AI Act) — Article 10