KAleep-Net: A Kolmogorov-Arnold Flash Attention Network for Sleep Stage Classification Using Single-Channel EEG With Explainability

Abstract

Sleep monitoring is essential for assessing sleep quality and understanding its broader implications for overall health. Although electroencephalography (EEG) remains the gold standard for sleep analysis, multichannel techniques are often cumbersome and impractical for real-world application. As a more feasible alternative, single-channel EEG offers greater practicality but still faces several persistent challenges, including reduced spatial resolution, feature instability, and limited clinical interpretability. To address these limitations, we propose KAleep-Net (Kolmogorov-Arnold based Sleep Network) for sleep stage classification. It employs a Multispectral Feature Pipeline to extract both fine-grained and coarse-grained features from single-channel EEG signals. It integrates a Temporal Sequencing Network with Flash Attention to capture rich and stable features effectively. The proposed approach achieved an accuracy of 86.5%, an F1-score of 79.6%, and a Cohen’s $\kappa $ of 79.9% on the Sleep-EDF-20 dataset, along with a 41.7% improvement in training speed. For the Sleep-EDF-78 dataset, it attained 85.0% accuracy, 77.0% F1-score, 78.0% $\kappa $ , and a 67.5% gain in training efficiency. On the SHHS dataset, the model achieved 86.4% accuracy, an F1-score of 0.79, and a $\kappa $ of 0.81, with an 8.18% improvement in training speed. For interpretability, an integrated gradient technique was adopted to enhance decision transparency and promote clinical adoption. The framework offers an efficient solution for sleep staging in resource-constrained environments with clinically trusted insights for single-channel EEG-based sleep monitoring.