On 21 September 2026 the Official Journal of the European Union published Commission Implementing Regulation (EU) 2026/2098 on the minimum metadata for datasets intended for secondary use under Regulation (EU) 2025/327, the EHDS (EUR-Lex), whose distinction between primary use, secondary use and prohibited uses we have already described. A few days earlier the obligation in Article 3(1) of Regulation (EU) 2023/2854, the Data Act, had become operational; Article 50 applies it to connected products “placed on the market after 12 September 2026” (EUR-Lex), as we explained with regard to access by default to data from connected products from 12 September 2026. The two deadlines are close in time, but the two acts have a different relationship with Artificial Intelligence: the EHDS names the training of algorithms among the purposes of secondary use and links data quality to Regulation (EU) 2024/1689, the AI Act; the Data Act is a horizontal regulation on access to and use of data, which does not regulate AI. We follow the path of a health dataset intended for training.
Training among the purposes of secondary use: Article 53 of the EHDS
Article 53(1) allows access for secondary use only if the processing is necessary for one of the purposes listed. Scientific research under point (e) includes “training, testing and evaluation of algorithms, including in medical devices, in vitro diagnostic medical devices, AI systems and digital health applications” (EUR-Lex). Recital 61 adds that activities related to research include “training of AI algorithms that could be used in healthcare or the care of natural persons”.
Training an AI system is a purpose named by the legislature, not an interpretative construction. It nonetheless remains a case of scientific research, subject to the prohibitions in Article 54. The data access application must explain the intended use in detail, describe the format and sources of the data, justify any request for pseudonymised data and contain “a description of the tools and computing resources needed for a secure processing environment” (Article 67(2)). Chapter IV applies from 26 March 2029, although the Regulation generally applies from 26 March 2027 (Article 105; see NicFab).
The secure processing environment: access, anonymisation and the prohibition of re-identification
Access takes place “only through a secure processing environment”, restricted to the persons listed in the data permit, with individual and unique user identities and logs of access “kept for at least one year” (Article 73(1)). Data are provided in anonymised form where the purpose allows it, and in pseudonymised form only if the user has sufficiently demonstrated that it cannot do without them (Article 66). Only “non-personal electronic health data, including electronic health data in an anonymised statistical format” may be downloaded (Article 73(2)); results “shall contain only anonymous data”, and re-identifying or attempting to re-identify the persons concerned is prohibited (Article 61) (EUR-Lex).
In the study on the patient’s digital twin put to the test of the GDPR, the AI Act and the EHDS we observed that a model trained on real data can memorise them and that cross-referencing with external sources, the so-called linkage attack, makes the risk of re-identification concrete. In the profile of the AI Security Specialist and of the attacks specific to AI systems we recalled that the NIST taxonomy includes attacks on the confidentiality of data and models and that the mature profile practises controlled offensive testing, including red teaming campaigns. In our view these are the competences needed to document, before requesting a download, whether the results make it possible to trace back to individuals: the prohibition of re-identification requires a technical check of what leaves the secure environment.
Data quality and bias: the express link between the EHDS and the AI Act
The express link is in Article 78. The data quality and utility label, optional in general and mandatory for datasets collected with public funding, covers “the completeness, uniqueness, accuracy, validity, timeliness and consistency of the data”, data quality management processes, “including review and audit processes, and bias examination”, and, where applicable, the “representativity of the population sampled”. Paragraph 6 entrusts the Commission, by 26 March 2027, with the technical specifications of the label and provides that those acts “shall take into account the requirements in Article 10 of Regulation (EU) 2024/1689” (EUR-Lex). Neither Article 78 nor Article 10 names sex; on the representativeness of health data by sex, see health data and artificial intelligence.
Article 10 of the AI Act subjects the training, validation and testing data sets of high-risk systems to governance practices that include the origin of the data and “examination in view of possible biases”, and requires them to be “relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose” (EUR-Lex). Regulation (EU) 2026/1744, the Digital Omnibus, deleted Article 10(5), placing in the new Article 4a the basis for processing special categories of data for the purposes of bias detection and correction, and set the application of the rules on high-risk systems at 2 December 2027 for Annex III and 2 August 2028 for Annex I (EUR-Lex). As Nicola Fabiano observes, Article 4a, inserted in Chapter I, already operates, while the Article 10 obligation it serves remains deferred; recital 9 justifies this by reference to preparation for compliance, and paragraph 2 specifies that the provision does not create any obligation to detect and correct bias (NicFab). In our view, for those training on pseudonymised health data that basis is added to the conditions of the EHDS and of Regulation (EU) 2016/679, the GDPR, without replacing them. The legislature has expressly linked the quality label for health data to the requirements of the AI Act. The label, however, describes a dataset, whereas Article 10 measures adequacy against the purpose of a specific system: in our view the label informs the assessment of the system’s provider, but does not replace it.
Implementing Regulation (EU) 2026/2098 requires metadata to be expressed “by using the definitions, structure, cardinalities and controlled vocabularies set out in the HealthDCAT-AP” and provides, among the elements, for the origin and generation process of the dataset, the estimated number of unique individuals, the temporal coverage, the variables and the type of dataset, including synthetic; it applies from 26 March 2029 (EUR-Lex). We consider these elements a useful starting point for documenting the origin of the data required by Article 10(2), point (b), of the AI Act.
On this ground the link with professional profiles is already written in our articles. In the profile of the AI Data Engineer and of quality that becomes compliance we identified systematic validation, completeness and consistency checks and traceability of origin as the point at which the craft meets the governance of training data sets required by the AI Act; in the portrait of the AI Data Scientist and of the method that makes the difference we linked to the same requirements the verification of the representativeness of data and the measurement of the effects of models on different groups. Article 78 translates those activities into elements of a public label.
The scope of the Data Act: access to and use of data from connected products
The Data Act contains no rules on the training of models: machine learning appears only in recital 33, where data intermediation services “could play an instrumental role in aggregating access to data so that big data analyses or machine learning can be facilitated”, provided that users remain in full control (EUR-Lex). For those intending to train models on data from connected products, what matters is who may use them: the data holder may use readily available non-personal data only “on the basis of a contract with the user” and does not make them available to third parties for purposes other than the performance of that contract (Article 4(13) and (14)); information derived from data by means of complex proprietary algorithms is excluded from the obligation to make data available (recital 15). The Data Act establishes who may use data from connected products, not what training data must be like.
The EHDS, for its part, includes among the categories for secondary use the data automatically generated by medical devices and data from wellness applications (Article 51(1)) and is without prejudice to the Data Act (Article 1(3)). In our view, for a medical device that is also a connected product, the Data Act governs access by the user, the EHDS governs secondary use and the AI Act, if the trained system is high-risk, governs the quality of the data sets.
Critical aspects: models leaving the environment, the expected implementing acts and the role of professional profiles
The first aspect concerns the trained model: the EHDS governs the download of non-personal data, but does not qualify the parameters of a model trained in the secure environment. We consider it prudent to treat their release as a download request, supported by tests of resistance to re-identification.
The second concerns implementing acts. Articles 73(5) and 78(6) require the Commission, by 26 March 2027, to lay down the requirements for secure environments, including “the technical characteristics and tools available to the health data user”, and the specifications of the label, applicable from 26 March 2029 (Article 105). In the search we carried out on EUR-Lex at the time of writing we did not find them.
The third concerns the addressees: these regulations are addressed to health data access bodies, data holders and users, and providers of AI systems, not to professional roles. The links with the profiles are our own assessments, limited to the cases in which our articles already connect the competence to the legal requirement. Certification of a person’s competences does not transfer the obligation from the organisation to the professional. For access by default under the Data Act we found no link of this kind: it is an obligation concerning product design.
Conclusions
For those intending to train, test or evaluate AI systems on health data, the EHDS offers from 26 March 2029 an express purpose, a mandatory place of processing and a language of quality built taking into account Article 10 of the AI Act. The obligations remain with health data access bodies, data holders and users, and the providers of the systems; the implementing acts on secure environments and on the label are missing; the Data Act remains the framework for access to data from connected products. We consider it useful to prepare now the documentation required by Article 67 of the EHDS and by Article 10 of the AI Act, entrusting to data engineering, data science and security competences the verification of quality on input and of non-re-identifiability on output: it is advisable to build checks that will stand up both to the EHDS authorisation and to the AI Act documentation.
Author: Valentina Grazia Sapuppo
