Kolmogorov–Arnold Networks: Comparative Analysis of Polynomial Bases for ECG Time-Series Forecasting

Document Type

Article

Publication Date

Spring 5-1-2026

Abstract

This study examines four Kolmogorov-Arnold Networks (KANs) variations where B-splines were replaced with polynomials as approximation functions—Chebyshev, Hermite, and Legendre—for  electrocardiogram forecasting on the MIT-BIH arrhythmia dataset. The existing architectures of KAN variants were surveyed and compared to the suggested models. In our research, all networks have a three-layer structure and are trained using different polynomial classes instead of B-splines to approximate the learnable activation functions. Our findings reveal an accuracy-efficiency trade-off: locally supported B-splines improve predictive precision, whereas globalpolynomial KANs have a better accuracy-to-cost ratio, making them appealing for resource-constrained time-series applications.

Share

COinS