Federated Learning with Differential Privacy for Imbalanced Clinical Data
A new study proposes a framework combining Federated Learning (FL) and Differential Privacy (DP) for cardiovascular risk prediction using imbalanced clinical data. The research addresses the trade-off between privacy and clinical utility by integrating SMOTETomek at the client level to handle class imbalance. Initial experiments with standard methods yielded zero recall, but the hybrid approach improved performance. The framework is designed for collaborative health research on decentralized data while providing formal privacy guarantees. The study is published on arXiv under ID 2508.10017.
Key facts
- Federated Learning enables model training on decentralized data while safeguarding patient privacy.
- Differential Privacy offers formal security guarantees when combined with FL.
- Integration of FL and DP introduces a trade-off between privacy and clinical utility.
- Severe class imbalance in medical datasets complicates the privacy-utility trade-off.
- The study implemented an FL framework for cardiovascular risk prediction.
- Initial experiments with standard methods resulted in a recall of zero due to imbalanced data.
- SMOTETomek (hybrid Synthetic Minority Over-sampling Technique with Tomek Links) was integrated at the client level.
- The research is published on arXiv with ID 2508.10017.
Entities
Institutions
- arXiv