Context card
Training data provenance is part of the device file, not a data-science record
Why does the origin of training data belong in the technical documentation of an AI-enabled medical device?
The short answer
Because the performance a device claims is only as valid as the data it was trained, tuned and tested on. The AI Act requires data governance records on origin, collection, preparation and examination for bias (Article 10, Annex IV), and FDA and its partner agencies ask that clinical study participants and data sets represent the intended patient population. Provenance links the claimed performance to the data behind it.
For: Data science, regulatory and documentation teams of AI-enabled medical devices
Key points
- Provenance answers: where does each data set come from, under which conditions was it collected, how was it prepared?
- Representativeness is checked against the intended purpose: population, setting, devices that produced the data.
- Training, tuning and test data must be kept apart, and the test data independent.
- Personal health data in training sets are subject to the GDPR; the file records the legal basis, not the data.
- A provenance record outlives the model: it is needed to explain a later performance drift.
The context
What provenance means here
Content provenance and data lineage describe where something comes from and how it was transformed. For training data of a medical device that means, per data set: the source (sites, registries, earlier studies), the time span, the acquisition devices and protocols, the inclusion and exclusion criteria, the labeling process and who labeled, the preparation steps and the split into training, tuning and test data. W3C PROV offers a vocabulary for such records.
Why it belongs in the file
Article 10 of the AI Act requires data governance for high-risk AI systems: design choices, collection processes and origin, preparation, assumptions, availability and suitability, examination for possible biases and measures to detect and correct them. Annex IV asks for the data requirements and data sheets in the technical documentation. The Good Machine Learning Practice principles of FDA, Health Canada and the MHRA (2021) ask that data sets represent the intended patient population and that training and test sets are independent.
Representativeness is judged against the intended purpose: a model trained on adult data does not support a claim for children; images from one scanner type do not support a claim for all scanners.
What the technical writer takes from it
The IFU states the population and the conditions the device was validated for, and the limits that follow from the data. Those statements must point back to the provenance record — otherwise the IFU makes a claim the file cannot support (see documenting model limitations).
What this card does not say
It does not decide which data are sufficient for a device. That is a question of the clinical evaluation and the notified body's assessment.
Questions readers ask next
- Do we have to put the training data into the technical file?
- No. The file holds the description of the data and their governance — origin, characteristics, preparation, splits, bias examination — not the data themselves.
- Is a public data set enough provenance?
- Its name and version are a start. The file also needs what the data set contains, how it was collected and labeled, and whether it represents the intended population.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) — EUR-Lex, July 12, 2024Law in force: Binding legislation that applies as of the review date.
- Good Machine Learning Practice for Medical Device Development: Guiding Principles — FDA, Health Canada and MHRA, October 27, 2021Final guidance: Guidance published in its final form by an authority or an official body. Guidance explains the law; it is not law itself.
- Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (draft guidance) — U.S. Food and Drug Administration, January 7, 2025Draft guidance: A draft for consultation. It may still change and must not be treated as the final text.
- Regulation (EU) 2016/679 (General Data Protection Regulation) — EUR-Lex, May 4, 2016Law in force: Binding legislation that applies as of the review date.
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.