Federated Learning in Practice Networks
FLIP-IT : Federated Learning in Practice Networks
Project Description
FLIP-IT investigates how federated learning can enable the collaborative development of medical AI models across distributed primary care practices without requiring patient records to be collected in a central repository. The project develops infrastructure for decentralized model training and federated analytics in outpatient healthcare.
By combining expertise in medical AI, federated learning, software infrastructure, and primary care, FLIP-IT aims to support the development and evaluation of AI models for clinically relevant risk prediction while addressing challenges related to privacy, interoperability, and distributed data.
- Funding: NEXT.IN.NRW, co-financed by the European Union through the EFRE/JTW Programme NRW 2021–2027
- Partners: Institute for Artificial Intelligence in Medicine (IKIM), docport GmbH, Flower Labs GmbH, Kassenärztliche Vereinigung Westfalen-Lippe, and Kassenärztliche Vereinigung Nordrhein
FLIP-IT investigates how machine-learning models can be trained collaboratively across distributed primary care data while patient records remain at the participating sites. The project studies how federated-learning approaches can be applied in heterogeneous healthcare environments with different local data sources and practice systems.
Kamp Lab contributes its expertise in federated learning and distributed machine learning to FLIP-IT. The group contributes to the development and evaluation of federated-learning methods for decentralized healthcare data.
Prof. Dr. Michael Kamp represents TU Dortmund University in the project with a focus on Federated Learning and AI.
Federated Learning and Privacy
FLIP-IT uses federated learning to enable collaborative model training across participating practices without requiring their local patient databases to be centralized. Model computations are performed locally, while information required for collaborative learning is exchanged through the federated infrastructure.
The project investigates privacy-enhancing techniques, including secure aggregation and differential privacy, to reduce risks associated with the exchange of model updates and aggregated information. Standardized data representations such as HL7 FHIR support interoperability between heterogeneous practice systems.
These techniques can reduce particular privacy risks but do not, by themselves, guarantee confidentiality or data-protection compliance. Their use therefore needs to be considered together with the technical, organizational, and legal framework of the system.
CKD Use Case
One of the clinical use cases considered in FLIP-IT is the early identification of complications such as chronic kidney disease (CKD). Predictive models can be evaluated in a federated setting in which data remain distributed across participating primary care practices.
The use case provides a practical setting for studying federated learning with heterogeneous clinical data and different local practice environments.
Expected Research Outputs
FLIP-IT aims to contribute methods and technical infrastructure for federated learning and federated analytics in primary care. Expected outcomes include:
- federated-learning methods for decentralized clinical risk prediction;
- approaches for working with heterogeneous data across participating practices;
- integration of privacy-enhancing techniques for distributed learning;
- interoperable interfaces for decentralized data processing;
- evaluation of federated-learning approaches in clinically relevant use cases.
The Federated Learning Approach
