TUDAMatch: Time-Series Unsupervised Domain Adaptation for Automatic Sleep Staging

Abstract

Automatic sleep staging based on deep learning approaches with Electroencephalography (EEG) can effectively assist experts in diagnosing personnel sleep disorders. The variance between the training data (i.e. source domain) and the practical data (i.e. target domain), and the cost of labeling has led to extensive studies on the unsupervised domain adaptation (UDA) method. However, most existing methods have limited performance on time-series data due to the lack of temporal dependency utilization, and training relies on the unavailable tremendous dataset. An unsupervised Time-series Domain Adaptation (TUDAMatch) is proposed to bridge the gap between the label distribution of source and target domains. In the data augmentation module, feature and consistency augmentation are designed to generate data as close to the true distribution as possible for better domain adaptation. The autoregressive discriminator module explicitly considers temporal dependence, and the domain adaptation module takes into account differences between source and target distributions. Experiments on the datasets of SLEEP-EDF, SHHS, and ISRUC-Sleep demonstrate the effectiveness of TUDAMatch, with an improved accuracy of 5.24%–13.79% and 2.85% compared with direct transfer and the reference [1], respectively. The TUDAMatch frame-work can promote the research of time-series unsupervised domain adaptation. The implementation code is available at: https://github.com/buptantEEG/TUDAMatch.