BD.20.003 – Managing Federated Data at Scale

AI can significantly improve our understanding of diseases when enough data is available and when reliable models can be built. Significant progress has been reported in this area. We are proposing to explore the setting when the approach will be scaled up to a fully operational situation. At least two problems need to be solved in that situation: 1) consent management: data comes from a large number of different sources that involve different ways of storing consent, under different jurisdictions, using different information models, etcetera; and 2) model management: models must be maintained and managed to ensure that they are still the right model for the right problem.

Consent management involves a (meta)data sharing framework that integrates different consent structures and requires semantic understanding of each patients’ consent form. AI model management requires a ‘DevOps’ type approach where constant and evolving evaluation of model performance and validity is needed.

These topics generate scientific questions, both fundamental questions and applied questions. For instance, when consent of one data subject is withdrawn, re-training is necessary: how to do this in situations where federated learning and/or multi-party computation techniques have been applied?

The proposal is to build upon insights that were obtained from for instance the NWA Startimpuls project VW Data (specifically, P7), the H2020 Big Medilytics (federated learning), PHT, and to explore how to make better use of the results and insights from IDS and GAIA-X.

Keywords

AI, consent management, data management, data sharing, health, machine learning, secure learning