With the increasing adoption of predictive healthcare systems, Federated Learning (FL) has emerged as a privacy-preserving paradigm for collaborative model training, which is crucial for sensitive health data. This study investigates the integration of Kolmogorov-Arnold Network (KAN) with FL to enhance predictive healthcare applications. KAN’s innovative implementation of spline-based activation functions offers enhanced flexibility, interoperability, and efficiency in capturing complex nonlinear data patterns with fewer parameters than traditional neural networks. When benchmarked against standard Multilayer Perceptrons (MLPs), KANs show significant performance improvements in both real-world and synthetic electrocardiogram (ECG) datasets. Notably, KAN’s adaptive activation functions, particularly quadratic splines with a grid size of 30, effectively capture the complexities of time-series data, achieving a test accuracy of 93.73% and an F1-score of 92.90% on the MIT-BIH arrhythmia dataset. Furthermore, KAN outperformed MLP on synthetic ECG data, highlighting its potential for broader medical applications. Despite its higher computational overhead, federated KAN shows considerable promise in healthcare settings that require both privacy and high accuracy in modeling complex data patterns. The findings underscore the need for optimization to address computational inefficiencies in KAN-based FL environments while establishing a foundation for future applications of KAN in time-series-based medical diagnostics.