Feature extraction and intelligent diagnosis of ECG signals based on KANs and xLSTM

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Cardiovascular disease (CVD) is a leading global cause of death.
• Accurate and timely diagnosis of cardiac arrhythmias is essential for CVD prevention and treatment.
• Existing ECG-based arrhythmia detection methods face challenges in achieving high accuracy.

Purpose of the Study:

• To propose and evaluate two novel deep learning architectures, Kolmogorov-Arnold Networks (KANs) and xLSTM, for enhanced arrhythmia detection from ECG signals.
• To compare the performance of KANs and xLSTM against existing deep learning methods using standard arrhythmia classification criteria.

Main Methods:

• Developed KANs with learnable spline activation functions and xLSTM with exponential gating and modified memory structures.
• Utilized focal loss function to address sample imbalance in ECG datasets.
• Classified arrhythmias based on Association for the Advancement of Medical Instrumentation (AAMI) standards.
• Evaluated models on the MIT-BIH (109,262 samples) and St. Petersburg INCART (166,909 samples) databases.

Main Results:

• Both KANs and xLSTM demonstrated superior performance in accuracy and F1 score for classifying five major arrhythmia categories.
• KANs achieved higher accuracy with lower computational complexity.
• xLSTM improved accuracy by approximately 0.41% compared to four popular deep learning methods.
• On the INCART database, KANs and xLSTM reached accuracies of ~96.87% and ~97.15%, respectively, showcasing versatility.

Conclusions:

• KANs and xLSTM represent significant advancements in deep learning for ECG-based arrhythmia detection.
• These models offer improved accuracy and efficiency, aiding in the prompt diagnosis and management of cardiovascular diseases.
• The demonstrated versatility across different databases highlights their potential for widespread clinical application.