Abstract
Sleep apnea is an increasingly prevalent and potentially serious sleep disorder marked by repeated interruptions in breathing during sleep. Polysomnography (PSG) serves as the clinical gold standard for diagnosis; however, its high cost, complexity, and limited accessibility hinder large-scale and home-based screening efforts. This study introduces a Heuristic Deep Learning (HeurDL) framework designed for the automated classification of sleep apnea events utilizing single-channel electroencephalogram (EEG) signals. The proposed framework combines wavelet-based EEG sub-band decomposition, heuristic domain-driven feature selection, and a lightweight one-dimensional convolutional neural network (1D-CNN) to enhance classification performance while minimizing computational complexity. EEG sub-bands of physiological significance are examined to identify distinguishing temporal, spectral, and nonlinear features linked to neural patterns associated with apnea. The proposed method differs from traditional end-to-end deep learning approaches by explicitly integrating heuristic knowledge from EEG physiology and empirical signal analysis, which improves interpretability and generalization. The framework has been implemented and assessed using publicly available benchmark EEG datasets, resulting in an overall classification accuracy of 91.2%, surpassing multiple existing EEG-based and wavelet–CNN hybrid methods. The findings indicate that heuristic-guided deep learning serves as an effective, scalable, and non-invasive approach for practical sleep apnea screening and decisionsupport applications.
Keywords: Convolutional Neural Network, Deep Learning, Electroencephalogram, Heuristic Learning, Sleep Apnea Detection.