Detection of K-Complexes and sleep spindles using a hybrid TCN and KAN architecture

Sleep disorders are a major public health concern. Electroencephalogram (EEG) analysis offers an effective approach for evaluating neurological status using K-complexes and sleep spindles as key biomarkers for sleep stage classification. This study proposes a hybrid architecture based on Temporal Convolutional Network (TCN) and Kolmogorov-Arnold Network (KAN). The model was evaluated on the DREAMS dataset and achieved a detection accuracy of 93.18% for K-complexes and 86.26% for spindles, outperforming baseline TCN models by 1%–2% across metrics. These results validate KAN's efficacy in time-domain signal classification. Notably, multi-layered KAN configurations failed to yield additional performance gains. Furthermore, we developed a sleep stage identification framework leveraging EEG biomarkers to consolidate physiologically similar stages into three macro-categories, attaining classification accuracies of 79.7% (K-complex subset) and 68.4% (spindle subset) on the DREAMs dataset. Extending this approach, we implemented a five-stage recognition system based on sleep phase duration ratios, which achieved 81.6% accuracy on the Haaglanden Medisch Centrum dataset. This work supports the feasibility of automated sleep staging using characteristic EEG waveforms.