EU AI Act Article 10 Health Data Compliance: A Practical Guide for Health AI Teams

The AI Act entered into force on 1 August 2024 and applies in stages. AI systems are classed as high-risk either because they are safety components of, or are themselves, products covered by EU harmonisation legislation listed in Annex I, which includes the Medical Devices Regulation and the In Vitro Diagnostic Medical Devices Regulation, or because they fall within the use cases listed in Annex III.

Most AI used for diagnosis, triage, treatment planning or clinical decision support is either a medical device or a component of one. Obligations for Annex III systems apply from 2 August 2026, and for Annex I products from 2 August 2027, subject to any amendments adopted by the EU institutions. Teams building models now are building them for that regime.

Article 10 requires that training, validation and testing datasets for high-risk AI are subject to data governance and management practices appropriate to the intended purpose. Those practices must address design choices, data collection processes and the origin of data, and, in the case of personal data, the original purpose of collection.

Datasets must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. Providers must examine datasets for possible biases likely to affect health and safety or fundamental rights, and take appropriate measures to detect, prevent and mitigate them.

• Documented provenance: where data came from and why it was originally collected.
• Documented preparation: annotation, labelling, cleaning, enrichment and aggregation steps.
• Assessed suitability: availability, quantity and suitability of the data for the purpose.
• Bias examination: a recorded process for identifying and mitigating bias, and for identifying data gaps.

Article 10(5) allows providers, exceptionally and to the extent strictly necessary for bias detection and correction, to process special categories of personal data, including health data. All of the following conditions must be met:

• Bias detection and correction cannot be effectively fulfilled by processing other data, including synthetic or anonymised data.
• The special category data is subject to technical limitations on re-use and to state-of-the-art security and privacy-preserving measures, including pseudonymisation.
• The data is secured and protected with suitable safeguards, including strict controls and documentation of access, so that only authorised persons with confidentiality obligations can reach it.
• The data is not transmitted, transferred or otherwise accessed by other parties.
• The data is deleted once the bias has been corrected or the retention period has ended, whichever comes first.
• Records of processing activities include the reasons why processing special category data was strictly necessary and why the objective could not be achieved with other data.

Article 10(5) does not replace GDPR. A lawful basis under Article 6 and a condition under Article 9 are still required for any processing of health data, and data protection impact assessments remain mandatory where processing is likely to result in high risk. The AI Act adds a second layer of documentation focused on data quality and bias rather than lawfulness.

Roles also differ. GDPR speaks of controllers and processors; the AI Act speaks of providers and deployers; the medical device rules speak of manufacturers. A single organisation may hold different roles under each framework, and its documentation must reflect all of them.

Article 10 is far easier to satisfy when the evidence is generated by the data pipeline itself. If every record entering a training set carries its consent basis, its origin and its permitted purposes, and if access to special category data is technically constrained and logged, the documentation writes itself.

Consenta provides that layer: purpose-specific, revocable consent evaluated at the point of use, federated access so data is analysed where it is held rather than transferred, de-identification with measurable k-anonymity, and an integrity-protected audit trail of every access. Consenta is in deployment and in discussion with organisations building and deploying health AI.