The PHEMS project (an EU Horizon initiative) is a groundbreaking effort poised to revolutionise healthcare data management and utilisation across Europe. It is particularly focused on addressing the challenges posed by privacy concerns and the complexity of data sharing due to varying interpretations of the EU General Data Protection Regulation (GDPR).

To address these systemic barriers, PHEMS is developing the Paediatric Health Data Space (PHDS): a trusted, privacy-preserving ecosystem dedicated to sharing paediatric health data. PHDS leverages the foundational technical infrastructure and governance framework established by PHEMS, thereby making the secondary use of health data ethical, efficient, and scalable. This capability will significantly advance research, benchmarking, and innovation in paediatric care.

My role involves leading one of the three clinical use-cases, which focusses on developing AI models to accurately predict drug exposure in patients suffering from rare bleeding disorders. A major hurdle in this project is the inherent sparsity of data for these patient cohorts, compounded by the fact that treatment is often administered at home and thus missing from the electronic health record (EHR), which serves as our primary data source within PHEMS.

To overcome this, we have developed a robust Bayesian generative model capable of reliably inferring missing dosage information based on readily available patient characteristics and lab measurements. By employing Bayesian methods, this model automatically incorporates prediction uncertainty, which we can naturally propagate to our drug exposure model to mitigate the risk of overfitting. Our model significantly reduced the Mean Absolute Percentage Error (MAPE) from a current standard of 30–50% down to approximately 20% when compared to true dosages (where available in clinical notes). Furthermore, the model’s robustness is demonstrated by a significant reduction in the quantile calibration error, dropping from 0.1–0.2 to 0.04. Finally, our complete model pipeline is designed for continuous improvement, allowing the generative model to be iteratively refined based on updates and insights from the drug exposure model.

As use-case leader, I am responsible for defining and delivering all promised project deliverables and ensuring the technical infrastructure meets our requirements. I worked with clinicians to complete our OMOP data dictionary so that partners operate using common vocabulary when referencing specific variables. I regularly meet with our technical staff to discuss the federated architecture, update upper management on our progress, and supervise (PhD) students working on the project. Technically, I manage the full deployment cycle on the Azure platform: setting up the federated nodes, preparing containerized model code, and overseeing the deployment of federated learning code.

Crucially, as we are the primary use-case for training end-to-end AI models in a traditional online federated learning framework, our project sets the standard for PHDS utility, especially in the context of rare disease. We have recently begun enrolling new centres into the project, and our use-case is an accessible entry points for new members owing to its relatively simple data mapping requirements.

The project is set to end in September, with the PHDS and the results of the project being presented at the Paediatric Innovation Day 2026 in Helsinki.