The European Health Data Space (EHDS) aims at transforming health care delivery, innovation, and research across Europe [1]. The EHDS is based on a legal framework on which the European Parliament and the Council reached a political agreement in spring 2024. It has the following three overarching goals: control over personal health data, developing a market for electronic health records (EHRs), and facilitating secondary data use.

This goal is to ensure that individuals can access and manage their own health data securely, which is at the core of the EHDS vision. By enabling portability of data across the European Union (EU), the initiative seeks to empower the citizens and health care providers to make better-informed decisions. Notably, member states can offer a complete drop-out to their citizens.

The EHDS seeks to initiate a competitive market for EHR systems that are accessible, efficient, secure, and interoperable across member states.

The EHDS aims to not only break down silos that hinder the flow of data between stakeholders in the health care system, but also promote the secondary use of these data for research, innovation, policy-making, and regulatory activities.

The EHDS primary legislation provides an overarching framework but does not provide any technical infrastructure or implementation details [2]. However, the secondary legislation and guidance documents from ongoing initiatives such as TEHDAS2 [3] are working on further details to support the implementation in the EU member states.

The ambition of the EHDS to establish a market for EHRs and to facilitate the secondary use of health data is challenging considering the current heterogeneous interpretation of the General Data Protection Regulation (GDPR) regarding the use of sensitive patient-level data [4]. While processing of health data without the explicit consent of patients is principally possible for research purposes within the GDPR, this typically requires careful documentation, risk assessments, and compliance agreements, all of which pose significant challenges to collaborative research initiatives. Moreover, ethical board approvals, based on detailed study protocols, are often necessary. Altogether, fulfilling all legal and ethical requirements can easily last 1 to‐2 years in practice.

Subsequently, access to data often further necessitates additional agreements defining the terms of, eg, costs, split of intellectual property rights, and liability, which require additional time, effort, and trust between parties. Moreover, national variations exist in the implementation of the GDPR and the requirements for ethical board approvals throughout the EU. Overall, the current legal landscape provides robust protection for patient data; however, its complexity complicates and slows down innovation projects, and problems multiply with the number of involved parties, sectors, and jurisdictions.

From a technical point of view, differences in data formats and standards are further hurdles to realize the requested portability of health data across the EU. Hence, initiatives like the Observational Medical Outcomes Partnership (OMOP) aim at establishing a common data model (CDM) designed to standardize the structure and content of observational health data [5]. In consequence, multiple initiatives (eg, EHDEN [6] and DARWIN-EU [7]) have been initiated to map real-world data collected during routine health care to OMOP and to develop the corresponding harmonization processes [8]. These initiatives thus address semantic interoperability of health data as a prerequisite for portability. However, data harmonization processes are not only complex, time consuming, and error prone, but also raise the question of how to deal with valuable non-standard data elements, which may only exist at national or institutional levels. The EHDS could amplify the impact of data harmonization initiatives through the adoption of standards and technical requirements across EU countries.

Decentralized data analytical approaches, including federated machine learning (FL) [9], offer a solution to overcome legal and technical challenges by enabling collaborative data analysis without necessitating centralized data storage. In FL, models are trained locally on data held by individual organizations. Only the model updates, not the raw data, are shared with a central server, which orchestrates the training process. FL enhances privacy by keeping sensitive data on-premises while enabling collaborative model training. Swarm learning (SL) [10] is a FL variant that removes the need for a central coordinating server. Instead, a blockchain is used to ensure direct and secure communication between local servers. SL offers improved scalability and faster model convergence compared to traditional FL, because it supports asynchronous learning and thus reduces the communication bottleneck with a central server. Major commercial players today provide robust libraries for federated data analysis, including descriptive summary statistics and FL/SL, thus lowering the barrier to employing such technologies in practice. Projects aiming for the implementation of EHDS-compatible platforms can thus build on relatively mature technology here.

Although decentralized data analysis reduces the need for data transfer, it is not immune to risks. For example, machine learning (ML) model updates can inadvertently reveal sensitive information through reverse engineering or gradient leakage [11]. Consequently, approaches such as differential privacy and encryption technologies remain essential safeguards. Differential privacy amounts to adding a defined amount of noise to model gradients during local model updates. However, this can degrade model performance and can only be mitigated by prolonged model training, which often proves computationally impractical.

Another challenge arises from the propagation of biases inherent in local datasets to the aggregated global model. This issue can be intensified when larger datasets have a disproportional influence. Furthermore, variations in data quality and heterogeneity across sites can affect ML model performance, necessitating rigorous validation and harmonization efforts across all parties participating in a decentralized data analysis. At the same time, the identification of possible reasons for biases in an ML model or generally poor prediction performance is extremely hard to detect if data scientists have no direct access to data.

While recent techniques allow the provision of human-understandable explanations of predictions produced by neural networks and can thus help to identify possibly model biases [12-14], these methods are not guaranteed to conform between models trained on different local and pooled datasets [11]. FL/SL approaches are thus in a certain tension with the general ambition of trustworthiness of AI formulated by the EU [15].

On a technical level, decentralized data analysis, including FL/SL, requires setting up a cloud computing infrastructure. For this purpose, multiple commercial platforms are now available. In addition, European projects such as IDERHA (Integration of Heterogeneous Data and Evidence towards Regulatory and HTA Acceptance [16]) aim for setting up a scalable cloud computing infrastructure with security measures that meet the requirements of the European health care sector. Still, there is a need to critically evaluate those solutions regarding computational infrastructure (eg, support of GPU usage), security of communication between organizations, authentication of users, and support of FL/SL. Notably, the latter may require opening specific ports at each of the participating sites, which in turn may only be possible on specific servers that are hosted in a demilitarized zone. Finally, interoperability is key. Hence, it is essential to map data at each participating organization to a standardized CDM such as OMOP.

The legal challenges outlined above necessitate that research projects address technical and legal data governance from the outset, as they are integral components of improving access to and secondary use of health data. Effective legal facilitation of data-driven health innovation must ensure robust protection of data subject rights while mitigating compliance risks for involved entities. Simultaneously, the legal governance framework should be designed to minimize overregulation and redundancy, as such issues directly undermine the workability and impact of the technical solutions it aims to support. Given the complexity of the current regulatory and technical landscape, a scalable and modular governance approach is essential. IDERHA has developed and successfully piloted such a data governance model, which has already been adopted by several other research initiatives, such as CERTAINTY [17].