Ensemble learning framework for robust sleep stage classification using single-channel EEG

Sleep stage classification accuracy often suffers from inter-subject variability and signal artifacts. This study presents a novel ensemble learning framework for robust sleep stage classification using single-channel EEG data from the Physionet database. We develop specialized base classifiers optimized for each sleep stage transition and combine their outputs using a stacking approach with a meta-learner. Our framework employs confidence-weighted voting and a novel error-correction mechanism that identifies and rectifies physiologically implausible sleep stage transitions. Results demonstrate that the ensemble approach achieves 91.3% accuracy, outperforming individual classifier performance by 4.7-7.2%. Notably, the framework shows significantly improved robustness to artifacts, maintaining 89.6% accuracy when tested on noisy segments that cause individual classifiers to fail. The error-correction mechanism successfully identifies 93.4% of physiologically implausible transitions, improving temporal consistency. This methodology provides a powerful approach for reliable sleep staging in home environments where recording conditions may be suboptimal, offering potential for improved sleep disorder diagnosis outside laboratory settings.