Cluster Domain-Aware Client Selection for Federated Learning in The Healthcare Field (CDCSF)

Client selection remains a critical challenge in Federated Learning (FL). Resource-aware strategies aim to reduce training delays and mitigate stragglers by selecting an appropriate subset of clients in each round. However, these methods prioritize computationally strong clients and exclude resource-constrained clients. In healthcare settings, this approach is impractical because it removes entire domains from training, which harms generalisation. To address these challenges, we propose CDCSF, a domain-aware client selection framework that re-partitions clients into domain-homogeneous groups in each iteration. CDCSF is a dynamic clustering framework based on the (EM) algorithm that clusters clients based on local feature prototypes to enhance domain diversity. The framework incorporates a reliability score derived from an exponential moving average of training time to favor efficient clients. Simultaneously, a fairness score is introduced to ensure that underrepresented clients can still contribute to the training. This approach preserves sufficient representation across all domains to improve model generalization and accelerate convergence. We conduct extensive experiments on a healthcare benchmark dataset to validate the effectiveness of CDCSF. The proposed method improves accuracy by 2% over FedAvg under domain shift and outperforms PoC by 8%. With the proposed adaptive client selection strategy, we further demonstrate that CDCSF converges significantly faster than baseline methods under heterogeneous resource and data conditions.