Objective: Automatic sleep staging models suffer from an inherent class imbalance problem (CIP), which hinders the classifiers from achieving a better performance. To address this issue, we systematically studied sleep electroencephalogram data augmentation (DA) approaches. Furthermore, we modified and transferred novel DA approaches from related research fields, yielding new efficient ways to enhance sleep datasets.
Approach: This study covers five DA methods, including repeating minority classes, morphological change, signal segmentation and recombination, dataset-to-dataset transfer, as well as generative adversarial network (GAN). We evaluated these mentioned DA methods by a sleep staging model on two datasets, the Montreal archive of sleep studies (MASS) and Sleep-EDF. We used a classification model with a typical convolutional neural network architecture to evaluate the effectiveness of the mentioned DA approaches. We also conducted a comprehensive analysis of these methods.
Main results: The classification results showed that DA methods, especially DA by GAN, significantly improved the total classification performance in comparison with the baseline. The improvement of accuracy, F1 score and Cohen Kappa coefficient range from 0.90% to 3.79%, 0.73% to 3.48%, 2.61% to 5.43% on MASS and 1.36% to 4.79%, 1.47% to 4.23%, 2.22% to 4.04% on Sleep-EDF, respectively. DA methods improved the classification performance in most cases, whereas the performance of class N1 showed a subtle degradation in the F1 scores.
Significance: Overall, our study proved that DA approaches are efficient in alleviating CIP lying in sleep staging tasks. Meanwhile, this study provided avenues for further improving the sleep staging accuracy using DA methods.