Unsupervised sleep staging system based on domain adaptation

Currently, most deep-learning-based sleep staging system relies heavily on a large number of labeled physiological signals. However, sleep-related data, such as polysommography (PSG), are often manually labeled by one or more than one professional experts with much effort. Meanwhile, due to physiological differences that existed among different subjects, how to boost the performance of trained models on an unseen dataset is still an open issue. One potential solution to this issue is to borrow knowledge from a labeled dataset to train an unlabeled or few labeled dataset by way of unsupervised or semi-unsupervised domain adaptation. To overcome the problem of insufficient labeled data for training robust sleep staging systems, this study aims to investigate the training of an unlabeled target sleep dataset from a labeled source sleep dataset in a deep learning framework, which integrates a conditional and collaborative adversarial domain adaptation module. To facilitate the network to learn domain-invariant features, a domain classifier is deployed for each feature extraction block at different scale. The input to the domain classifier at different level is the multilinear mapping of the sleep stage prediction vector and the corresponding feature vector at this level. It is assumed that the feedback of the class information provided by the network into the domain classifier can be beneficial to help the network to reduce the feature distribution distance between different domains. Experiments on public Sleep-EDF dataset demonstrate the effectiveness of the proposed approach. Compared to other domain adaptation approaches, the proposed approaches can provide better sleep staging performance in different model transferring tasks.

Automated summary

Most deep-learning-based sleep staging systems rely on labeled physiological signals, but manual labeling is time-consuming. Physiological differences among subjects make model generalization challenging, and training unlabeled data through domain adaptation can effectively address data insufficiency.