Abstract Introduction Automated sleep staging using deep learning is increasingly important for improving the efficiency and reproducibility of polysomnography (PSG) scoring. However, model performance can be constrained by the variability and limited diversity of available training data. Data augmentation offers a potential solution by artificially expanding the range of signal characteristics encountered during training, yet its impact on sleep staging performance remains underexplored. This study investigates whether applying augmentation techniques to spectrogram representations of PSG signals—including electroencephalogram, electrooculogram, and chin electromyogram—can enhance model accuracy and generalization. Methods Model inputs were sequences of time–frequency spectrograms computed from PSG channels recorded according to AASM standards. During training, each sequence was randomly augmented with one of five strategies: (1) no augmentation, (2) time masking, (3) frequency masking, (4) amplitude scaling, or (5) Gaussian noise. A deep learning architecture combining convolutional residual blocks and bidirectional LSTM layers was trained using full-night PSG recordings from 2,269 patients (36.71% female, 63.29% male; mean age 47.87 ± 17.48 years) collected as part of a large clinical sleep study. Model evaluation was conducted on the ANPHY-Sleep open dataset (N = 29), which is entirely independent from the training cohort and differs in demographic characteristics, collection site, and recording conditions. This separation provides a stringent assessment of model generalizability and robustness across populations. Models trained with and without augmentation were directly compared under this evaluation framework. Results Introducing augmentation strategies during training yielded measurable improvements in sleep stage classification. The augmented model achieved higher overall accuracy (84.0% vs. 82.6%) and a higher macro F1-score (0.789 vs. 0.779) compared with the non-augmented baseline. Improvements were observed across a majority of recordings, indicating that augmentation helped the model better accommodate natural variability in PSG signals. Conclusion This study demonstrates that targeted augmentation strategies can meaningfully enhance spectrogram-based deep learning models for automated sleep staging. By increasing the diversity of training inputs, augmentation improved performance on an external dataset and strengthened model resilience to inter-subject and recording variability. These findings support incorporating augmentation techniques as a practical approach for advancing clinical sleep staging systems. Support (if any)
Kaongoen et al. (Fri,) studied this question.