Federated Mayfly Optimization for Privacy-Preserving Chronic Kidney Disease Diagnosis

Chronic kidney disease (CKD) is a globally pervasive and insidious health crisis, frequently escaping detection until its advanced stages, which critically narrows the window for effective treatment and drastically worsens patient prognosis. Crucially, early and precise diagnosis is the key to enhancing patient outcomes and significantly cutting healthcare expenditures. While machine learning (ML) models offer remarkable potential for predicting CKD, their clinical adoption is hampered by the need for centralized datasets, which inevitably triggers major concerns regarding patient privacy and data security. Federated learning (FL) directly tackles this privacy dilemma by allowing multiple institutions to collaboratively train a global model without ever sharing their raw, sensitive patient data. Nevertheless, standard FL approaches, such as federated averaging (FedAvg), are known to suffer from poor performance and sluggish convergence, especially when dealing with the heterogeneous (non-IID) data distributions typical of real-world clinical environments. To overcome these performance bottlenecks, we introduce the Mayfly-based federated learning (MBFL) framework that embeds the Mayfly optimization algorithm (MOA) into the FL aggregation process. MBFL fundamentally improves both the convergence speed and the overall robustness of the global model across diverse data sources. We conducted a rigorous comparative trial using public CKD datasets, benchmarking the performance of MBFL against the foundational FedAvg and two other leading metaheuristic FL variants: Federated Particle Swarm Optimization (FedPSO) and Federated Sand Cat Swarm Optimization (FedSCSO). The results mark a definitive paradigm shift: MBFL achieved a remarkable classification accuracy of 99.2%, decisively outperforming all comparison algorithms. This unprecedented performance confirms MBFL as a performance leader in both data-independent (IID) and non-independent (non-IID) distributed scenarios. Ultimately, MBFL offers a streamlined, efficient, and collaborative new standard for CKD detection, far surpassing the capabilities of current models.