AI in healthcare depends on access to relevant data. Clinical records, diagnostic images and treatment histories can support the development of useful models, yet they also contain some of the most sensitive information an organisation can hold.
Access to relevant medical data is one of the main barriers facing healthcare innovators. AI developers need sufficiently large, well-prepared datasets to test whether their models work. Start-ups and SMEs often lack the institutional access, legal knowledge and infrastructure required to obtain them. Hospitals and public authorities, meanwhile, must protect patient privacy, comply with regulation and prevent security risks or inappropriate secondary use.
AI4Health.Cro addressed these needs through a structured model that connected safe data access with clearly defined healthcare challenges. The project treated medical data as a resource for responsible innovation, governed through institutional procedures, anonymisation, secure processing and expert oversight.
The challenge
Technical expertise alone is rarely enough to bring an AI solution into healthcare. Innovators must understand the clinical problem, identify the right data, establish legal access, prepare the dataset and demonstrate that results can be interpreted in a medical context.
For clinicians and healthcare institutions, the questions are equally demanding. They need to know who can access the data, for what purpose, under which conditions and with what safeguards. They also need confidence that proposed solutions respond to genuine clinical or public-health needs.
AI4Health.Cro created a pathway that addressed these requirements together.
The best-practice approach
At the beginning of the project, the consortium established a dedicated data group involving 10 of its 16 partners. It brought together expertise in data acquisition, data management, anonymisation, processing, modelling, healthcare practice and regulation.
The group developed a regulatory sandbox model referred to as a Secure Processing Environment. Its purpose was to enable real-world medical data to be extracted, anonymised, prepared and used under controlled conditions.
The work applied procedures developed within the European Health Data Space framework, with key involvement from the Croatian Institute of Public Health, the Croatian Health Insurance Fund and the Ministry of Health.
The process followed a series of practical steps:
- defining the healthcare problem,
- identifying the required data,
- securing access through appropriate institutional channels,
- preparing and anonymising the data,
- defining rules for secondary use,
- providing computing resources,
- and supporting users with clinical, technical and regulatory expertise.
The Ruđer Bošković Institute also compiled a catalogue of publicly available datasets suitable for AI development. The catalogue explained access conditions for different user groups and provided an alternative starting point where direct access to proprietary health data was unnecessary or unavailable.
Testing the model through innovation challenges
AI4Health.Cro applied this approach through three annual innovation challenges based on real healthcare problems. Each challenge gave participants a defined task, relevant data or data-based problem setting, a structured development period and access to expert support.
Predicting hospital readmissions
The 2024 challenge focused on predicting whether a patient would be readmitted to hospital within 30 days of discharge. It attracted 99 applicants organised into 27 teams.
Participants were asked to identify patterns associated with early readmission, build and interpret a predictive model, and develop a prototype interface capable of explaining individual predictions. The task was relevant to patient follow-up as well as hospital resource planning.
Classifying breast lesions
The 2025 challenge addressed the classification of suspicious lesions in mammography images. A total of 106 competitors applied through 34 teams, with 14 teams selected to continue developing their solutions.
Participants worked on predicting BIRADS scores, identifying suspicious lesions and creating prototypes of radiological decision-support systems. The challenge highlighted the specific requirements of medical imaging, including carefully labelled data, model interpretability and integration into an existing diagnostic process.
Identifying risks in diabetes care
The 2026 challenge, titled AI in the Service of Diabetes, recorded the highest participation of the three editions. It attracted 150 participants in 44 teams. Following the first evaluation, 24 teams comprising 85 competitors advanced to the development stage.
The task focused on type 2 diabetes and the risk that patients would stop following their prescribed treatment. Teams developed models intended to help healthcare professionals identify signs of medication non-adherence earlier.
The challenge was supported by AstraZeneca and Novo Nordisk Hrvatska. Insulogic won the competition with a platform designed to identify patients at increased risk and support clinical follow-up.
Benefits for different user groups
For start-ups and SMEs, the model reduced development risk. Teams could test ideas against real healthcare needs before investing further resources.
For researchers, students and developers, the challenges provided an entry point into health innovation and an opportunity to form multidisciplinary teams.
For clinicians, the model created a structured way to contribute medical knowledge, define relevant problems and assess whether proposed outputs could support practice.
For healthcare and public-sector organisations, it provided a controlled setting in which to explore the secondary use of health data while maintaining privacy, security and institutional oversight.
For the EDIH network, the approach offers a clear example of how project activities, infrastructure and user services can be connected to measurable participation, practical experimentation and ecosystem development.
Evidence and lessons learned
Participation increased from 99 applicants in 2024 to 150 participants in 2026. Across the three challenges, teams worked on early hospital readmission, breast cancer diagnostics and medication adherence in type 2 diabetes.
The experience showed that access to data is only one part of the process. Users also need a clearly defined problem, realistic tasks, mentoring, computing capacity, interpretation requirements and a route towards clinical validation.
These lessons strengthened AI4Health.Cro’s wider Test Before Invest services, including secure data access, clinical feedback, model interpretation and expert mentoring.
Transferable value
The AI4Health.Cro model is relevant to other European Digital Innovation Hubs, health projects and organisations working in regulated sectors.
Its central lesson is practical: health data can support innovation when access is structured, responsibilities are clear and technical development remains connected to real clinical needs.
By combining secure processing, institutional governance, expert support and challenge-based innovation, AI4Health.Cro turned a major barrier into a repeatable pathway from healthcare problem to tested solution concept.DIH