Project R-16884

Personalized Federated Learning for Heterogeneous Health Data Networks: From Methodological Development to Industrial Application in Pharmaceutical Research (Research)

Health data collected during routine care can help answer important questions about disease progression, treatment safety, and treatment effectiveness. Yet this information is spread across electronic health records, insurance claims, disease registries, and other real world data sources. Patient level records usually cannot be brought together in one central location because of privacy, governance, legal, and technical constraints. These sources also represent different patient populations and healthcare practices. A prediction approach developed using one source may therefore be unreliable in another, especially when a source contains relatively few patients or when the outcome of interest is uncommon. This project will develop personalized federated learning methods that support joint analysis across separate health data sources while the underlying patient records remain where they are held. The methods will combine patterns found across the wider network with information specific to each source. The research will focus on three practical challenges: adapting predictions efficiently to local populations, improving learning when outcomes or treatment groups are uncommon, and preventing useful shared information from being lost during local adaptation. The methods will first be developed and evaluated using routine clinical data from an international multiple sclerosis network. The task is to predict whether a person's disability will progress within two years. Methods that meet predefined performance criteria will then be evaluated in pharmaceutical real world evidence research. In this setting, they will estimate the probability that a person receives one treatment rather than another based on their health information before treatment. These estimates are used to balance treatment groups, helping researchers make more credible comparisons of treatment safety and effectiveness when random treatment assignment is not possible. The project will deliver a validated analytical framework, open source software, practical documentation, scientific publications, and a workflow tested in an industrial setting. By making more reliable use of existing health data, it aims to support earlier evidence generation for small patient populations, uncommon outcomes, and newly introduced treatments.

01 September 2026 - 31 August 2027