Feature extraction and intelligent diagnosis of ECG signals based on KANs and xLSTM.
Journal:
Biosensors & bioelectronics
Published Date:
Oct 15, 2025
Abstract
Cardiovascular disease (CVD) is the top cause of mortality globally, making it crucial to diagnose arrhythmias promptly and accurately for the early prevention and treatment of CVD. While numerous methods exist for detecting arrhythmias using ECG signals, achieving more accurate detection remains a significant challenge. This paper proposes two novel deep learning architectures: Kolmogorov-Arnold networks (KANs) and xLSTM, a variant of long short-term memory (LSTM) networks. They are used to extract features from ECG signals to classify types of arrhythmias. KANs integrate learnable activation functions at the connections, replacing each weight parameter with spline function. xLSTM introduces an exponential gating mechanism and modifies the memory structure of the LSTM to create a sLSTM and a fully parallelizable mLSTM. In this study, we utilize focal loss function to resolve sample imbalance and the classification criteria are based on the standards set by the Association for the Advancement of Medical Instrumentation (AAMI) for arrhythmias. In classifying five major categories of 109,262 heartbeat samples in the MIT-BIH database, the two proposed architectures have advantages in metrics such as accuracy, F1 score. The KANs have lower computational complexity while maintaining a higher accuracy. The xLSTM improves the accuracy by about 0.41% compared to the four existing popular deep learning methods. In addition, on the St. Petersburg INCART database with 166,909 heartbeat samples, the two architectures achieve accuracies of approximately 96.87% and 97.15%, fully demonstrating their versatility on different databases.
