Abnormality Detection in Time-Series Bio-Signals using Kolmogorov-Arnold Networks for Resource-Constrained Devices

Abstract

AbstractThis study uses Kolmogorov-Arnold Networks (KANs) to analyze electrocardiogram (ECG) signals in order to detect cardiac abnormalities. These novel networks have demonstrated potential for application in biosignal analysis, particularly ECG, due to their flexibility and smaller parameter requirements, making them candidates for wearable devices. The network structure comprises a simple KAN model with a single hidden layer of 64 neurons. It was trained on the Telehealth Network of Minas Gerais (TNMG) dataset and tested for generalization on the Chinese Physiological Signal Challenge 2018 (CPSC) dataset. The KAN model delivered reasonably promising results, achieving an F1-score of 0.75 and an AUROC of 0.95 on the TNMG dataset. During the out-of-sample generalization test on the CPSC dataset, it achieved an F1-score of 0.62 and an AUROC of 0.84. It has also shown resistance to missing data channels by maintaining a reasonable performance, down to only a single lead left of ECG data instead of the initial 12 leads. Compared with traditional Multi-Layer Perceptrons (MLP) and Neural Circuit Policy (NCP, aka. Liquid Time Constant Networks), KANs exhibit superior flexibility, adaptability, interpretability, and efficiency. Their compact size and reduced computational requirements make them potential candidates for deployment on hardware, particularly in personalized medical devices.