Robust feature selection for improved sleep stage classification

Effective sleep stage classification requires identifying discriminative EEG features that remain consistent across different subjects. This study proposes an ensemble feature selection framework for robust sleep stage classification using the Physionet EEG dataset. We extract 40+ features from time and frequency domains, then employ multiple selection techniques including mutual information, recursive feature elimination, and Lasso regularization. Our ensemble approach ranks features based on selection frequency across methods and cross-validation folds, identifying a minimal effective feature set. Results show that our selected 12-feature subset achieves 95.6% of the performance of the full feature set while reducing computational complexity by 68%. The most discriminative features were spectral edge frequency, delta-band power, and sample entropy, which align with known neurophysiological sleep markers. Subject-independent validation confirms that these features remain consistent across individuals, with 85% overlap in top-ranked features. This robust feature selection methodology enables more efficient sleep stage classification algorithms and provides insights into the fundamental EEG characteristics that define different sleep stages.