Abnormality Detection in Time-Series Bio-Signals using Kolmogorov-Arnold Networks (KANs) Based Models for Resource-Constrained Devices

Abstract

Abstract This study investigates Kolmogorov–Arnold Networks (KANs) for biosignal analysis, using electrocardiogram signals as a case study. KANs provide flexibility and require few parameters, making them suitable for wearable and edge devices. A simple KAN model with one hidden layer of 64 neurons was trained on the TNMG dataset and tested on the CPSC 2018 dataset, achieving an F1-score of 0.75 and AUROC of 0.95 on TNMG, and an F1-score of 0.62 and AUROC of 0.84 on CPSC. The model also showed robustness to missing channels, maintaining reasonable performance with only a single ECG lead. Compared with traditional Multi-Layer Perceptrons (MLPs) and Neural Circuit Policies (NCPs), KANs demonstrated greater flexibility, adaptability, interpretability, and efficiency. Additionally, a shallow network (CKAN) that integrates a single Conv2dLSTM layer with a small set of KAN neurons, mirroring two architectures built with different NCP neurons for TinyML, achieved an F1-score of 0.84 and an AUROC of 0.97 on TNMG, and an F1-score of 0.72 and an AUROC of 0.92 on CPSC. Incorporating learnable sparsity, a key feature of NCP neurons, into KAN neurons surprisingly enhanced both performance and generalization. Even after pruning sparse weights, the model maintained strong performance, surpassing the counterpart without sparsity.