Data access and sharing are central to medical research, but such sharing must be accomplished in a manner that respects regulatory frameworks, particularly when patient privacy is concerned. Traditional biomedical data types and study designs have long shaped existing solutions; however, there are numerous factors that have limited the use of such solutions moving forward. First, recent developments in novel data generation mechanisms (e.g., wearable devices measuring ECGs, functional and diffusion MRI, retina scans, and speech patterns) and advances in data-driven artificial intelligence (AI) have created new complexities. Second, health emergencies, like the COVID-19 pandemic, have demonstrated the need for rapid and global research collaboration, further complicating the reuse of existing strategies. Third, there are increasingly blurred lines between biomedical research, clinical practice, and public health applications, raising additional challenges around data reuse and patient autonomy.

This Dagstuhl Seminar, situated at the intersection of computer science, medical research, and law, will explore the state of the art and the future of computational approaches to address these emerging challenges, structured around, but not limited to, four central topics:

1. Rapid data sharing: To protect privacy, medical research is shifting from sharing individual-level data towards more sophisticated mechanisms. Examples include federated learning, where models are based on statistical representations of the data only, AI-generated synthetic data that mimics the statistical properties of the original data, homomorphic encryption, secure multiparty cryptography, and secure enclaves. However, each of these techniques has strengths and weaknesses, offering opportunities for improvements and combinations. Moreover, it remains unclear which technique is most feasible for each use case and data type, particularly in emergencies requiring rapid responses.
2. Cross-border data flows: Processing and sharing health data across countries with varying legal and regulatory frameworks remain particularly challenging. Current technical solutions for privacy-preserving data sharing and processing often fail to address the need to share large datasets of local and regional cohorts across national borders. Consequently, there is a significant need for alignment between technical implementation options and legal requirements and innovative technical solutions that better balance data privacy and research interests.
3. Privacy-preserving AI: AI methods have the potential to revolutionize medicine, as seen, for example, in clinical decision support, medical imaging, and protein design. To enable medical AI applications to offer clinical decision support suitable for precision medicine, even larger amounts of multi-modal healthcare data must be processed in a privacy-preserving manner. Statistical disclosure control mechanisms, such as differential privacy, are a promising class of techniques in this context. However, how to optimally balance privacy protection and the performance of different AI models remains an open question, particularly when applied to modern data types like time series from wearable sensors, imaging data, or genomic or proteomic profiles.
4. Robustness across populations: Numerous studies show that privacy-enhancing technologies and healthcare AI models may not perform consistently across all patient groups, often due to gaps or imbalances in data. Data sharing technologies, mixing in synthetic data to enhance performance across different populations, or transfer learning methods offer potential strategies for improvement. However, many open questions remain regarding when and how these approaches can be applied, and under which circumstances it may be permissible to utilize algorithms with known performance variations.

We hope to spur a fruitful exchange allowing the involved communities to learn the requirements, problems, and solution strategies from each other, thus advancing the discussed technologies as well as their application and the understanding of their embedding into the wider societal framework.

Creative Commons BY 4.0

Murat Kantarcioglu, Bradley Malin, Fruzsina Molnar-Gabor, Fabian Prasser, and Daniel Rückert