Sleep Stage Classification Using EEG Signals and Visibility Graph Analysis: A Novel Approach

Abstract

Accurate classification of sleep stages is critical for diagnosing sleep disorders and monitoring neurological health. Traditional computerized sleep staging relies on spectral transformations, such as the Fast Fourier Transform (FFT) and Continuous Wavelet Transform (CWT), which frequently fail to capture the non-stationary, chaotic, and non-linear properties intrinsic to human brain activity. This study introduces an automated framework that maps single-channel electroencephalography (EEG) time series into complex network topologies via Visibility Graph Analysis (VGA). By evaluating geometric lines of sight between sequential sample amplitudes, visibility graphs preserve dynamical invariants and scaling properties of electrophysiological records. From each 30-second epoch graph, two non-linear descriptors are extracted: Maximum-Minimum Distance (MMD) and EnergySis (Esis). These representations are classified using an optimized 1D Convolutional Neural Network (CNN). Evaluated on the benchmark PhysioNet Sleep-EDF database, the proposed pipeline achieves an overall accuracy of 93.13%, sensitivity of 89.06%, specificity of 98.61%, precision of 90.45%, and an F1-score of 89.75%. Granular evaluation demonstrates an accuracy of 98.61% for Slow-Wave Sleep (N3) and 89.10% for Rapid Eye Movement (REM), while maintaining robust distinction between REM and Non-REM stages. These findings highlight the viability of graph-based non-linear dynamics for clinical polysomnography and embedded wearable neurological devices.