Deep Learning-Based EEG Frequency Domain Analysis for Classification of Disorders of Consciousness

Abstract

Disorders of Consciousness (DOC), including Minimally Conscious State (MCS) and Unresponsive Wakefulness Syndrome (UWS), are neurological conditions characterized by severe impairment in cognition and perception due to brain injury or degeneration. Traditional behavioral and imaging-based diagnosis methods are costly, subjective, and dependent on expert interpretation. To address these limitations, this study proposes a deep learning-based method that leverages EEG frequency domain features for objective DOC classification. Power Spectral Density (PSD) is extracted from EEG signals using the Welch method to capture multi-band neural oscillation characteristics. A two-dimensional Convolutional Neural Network (CNN) is then employed to automatically learn discriminative representations from the spectral maps. Dropout regularization and ReLU activation functions are integrated to enhance generalization and nonlinear feature extraction. Experimental evaluation on EEG datasets from MCS and UWS patients demonstrates that the proposed method achieves an average classification accuracy of 86.22%, significantly outperforming traditional machine learning approaches. The results confirm that PSD-based spectral encoding enhances robustness and interpretability by emphasizing frequency bands most associated with consciousness states. This framework provides an efficient, scalable, and objective solution for DOC classification, offering valuable support for clinical prognosis assessment and intelligent diagnosis.