Time-series augmentation methods for improved sleep stage classification robustness

Data augmentation can address limited training data and class imbalance in sleep stage classification. This study presents a comprehensive framework of EEG-specific augmentation techniques to improve model robustness using the Physionet dataset. We implement traditional time-series transformations (time warping, magnitude scaling, jittering) alongside novel EEG-specific augmentations that preserve sleep stage characteristics. Generative models including VAEs and GANs with spectral constraints are trained to synthesize realistic sleep EEG data. Our consistency regularization framework ensures models produce stable predictions for original and augmented versions of the same segment. Results show that augmentation improves overall classification accuracy by 5.8%, with particularly significant gains for underrepresented stages (8.7% for S1, 7.3% for REM). The curriculum-based augmentation strategy, which progressively increases transformation complexity during training, further improves robustness to signal quality variations. Expert evaluation confirms that synthetically generated EEG signals maintain the physiological characteristics of each sleep stage. This augmentation methodology enables more effective model training with limited data and enhances performance under challenging recording conditions.