Leveraging biased FedSDS as a federated learning strategy for research on non-IID multi-disorder studies

Abstract Federated learning is a powerful approach to overcome privacy and data sharing issues, especially in small sample healthcare settings. However, parameter aggregation strategies, such as FedAvg, are prone to fail in case of non-IID data distributions, which often arise in real-world scenarios. In fact, data provided by different institutions can exhibit systematic heterogeneity across cohorts, such as age, genotype, and life expectancy, reflecting institution-specific biases. Federated Synthetic Data Sharing (FedSDS) represents a valuable solution to the problem of non-IID data. It is a federated learning strategy based on locally generating synthetic datasets from real data and sharing them between the federated clients, avoiding the transfer of real sensitive data. FedSDS is applied to a survival prediction task, performed using a DeepSurv model architecture in 3 different healthcare use cases: Myelodysplastic Syndrome, Low Grade Glioma, and Breast Cancer. We demonstrate that this approach outperforms FedAvg and other strategies based on model parameters optimization in non-IID settings. In particular, the possibility to locally bias shared synthetic data allows preserving the local data distribution of each participating client and significantly improving the performances of federated models. This study shows that biased FedSDS is effective in three different healthcare use cases, such as Myelodysplastic Syndromes, Low Grade Glioma, and Breast Cancer. Our findings suggest that FedSDS may have potential relevance for healthcare applications, as it can support the selection and biasing of samples for dataset augmentation and improve the performance of federated models in non-IID multi-disorder settings. The potential use of this approach for clinical trial cohort selection, however, would require further investigation and dedicated validation.