From high-dimensional EEG to robust sleep deprivation detection via

초록

Sleep deprivation poses significant risks to cognitive function and physiological well-being, underscoring the need for accurate and automated detection systems. This study proposes a comprehensive EEG-based framework for classifying sleep-deprived states using optimized feature selection techniques. EEG recordings were acquired from 71 participants under both normal sleep and sleep-deprived conditions. Following extensive preprocessing, including time cropping, artifact removal via Independent Component Analysis, and bandpass filtering, features were extracted across theta, alpha, and beta frequency bands, along with nonlinear descriptors such as R & eacute;nyi entropy, correlation dimension, and Lempel-Ziv complexity, resulting in a 732-dimensional feature space. To enhance classification performance and reduce redundancy, three feature selection methods were employed: Spectral Discriminative Manifold Projection (SDMP), Laplacian Score, and Minimum Redundancy Maximum Relevance (mRMR). A comparative evaluation using multiple classifiers-Random Forest, k-Nearest Neighbors, Decision Tree, Support Vector Machine, and XGBoost-revealed that the combination of Laplacian score and XGBoost achieved the highest accuracy (98.16%) and the kappa score (0.9566), outperforming other configurations. These results indicate the potential of EEG-based models to support objective identification of sleep-deprived states in practical settings where timely fatigue detection is critical. By reducing redundancy in high-dimensional EEG features, the proposed framework can facilitate scalable deployment for clinical screening and safety-critical monitoring applications.

키워드