SimpleKANSleepNet: a Kolmogorov-Arnold network based sleep stage classification method.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Computational Neuroscience

Background:

• Automatic sleep stage classification is crucial for diagnosing sleep disorders.
• Existing methods like Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs) have limitations in flexibility and performance.
• Multilayer Perceptrons (MLPs) can be enhanced with learnable activation functions for improved modeling.

Purpose of the Study:

• To propose and evaluate a novel Kolmogorov-Arnold Network (KAN) based model for automatic sleep stage classification.
• To assess the performance of the SimpleKANSleepNet model using temporal and frequency features from multiple biosignals.
• To compare the KAN model's effectiveness against established CNN and GCN approaches.

Main Methods:

• A dual-stream CNN architecture was employed to extract temporal and frequency features from EEG, EMG, EOG, and ECG signals.
• A redefined Multilayer Perceptron (MLP) architecture utilizing Kolmogorov-Arnold Networks (KANs) with learnable activation functions was developed.
• The SimpleKANSleepNet model was evaluated on the ISRUC-S1 and Sleep-EDF-153 datasets.

Main Results:

• The SimpleKANSleepNet model achieved high performance metrics, including overall accuracy, F1-score, and Cohen's kappa on both datasets.
• Comparative analysis showed the KAN model outperformed existing CNN-based and GCN methods.
• Data balancing techniques and factor analysis were explored to further optimize classification performance.

Conclusions:

• The proposed KAN-based SimpleKANSleepNet model demonstrates significant potential for accurate and generalizable automatic sleep stage classification.
• KANs offer a flexible and effective alternative to traditional activation functions in deep learning models for biomedical signal processing.
• The study highlights the efficacy of KANs in handling complex physiological data for sleep analysis.