Journal
METHODS
Volume 204, Issue -, Pages 84-91Publisher
ACADEMIC PRESS INC ELSEVIER SCIENCE
DOI: 10.1016/j.ymeth.2022.03.013
Keywords
Pediatric sleep staging; EEG channel; Adversarial learning; Semi-supervised
Funding
- National Natural Science Foundation of China [62172340]
- Natural Science Foundation of Chongqing [cstc2021jcyj-msxmX0041]
- Fundamental Research Funds for the Cen-tral Universities [SWU020008]
- Young and Middle-aged Senior Medical Talents Studio of Chongqing [ZQNYXGDRCGZS2021002]
Ask authors/readers for more resources
This paper proposes a Bi-Stream Adversarial Learning network (BiSALnet) for pediatric sleep staging, which generates pseudo-labels with higher confidence through adversarial learning and semi-supervised methods. The model combines symmetric positive definite manifold structure and attention feature fusion module to improve the performance on local pediatric dataset and a well-known public dataset.
Despite the progress recently made towards automatic sleep staging for adults, children have complicated sleep structures that require attention to the pediatric sleep staging. Semi-supervised learning, i.e., training networks with both labeled and unlabeled data, greatly reduces the burden of epoch-by-epoch annotation for physicians. However, the inherent class-imbalance problem in sleep staging task undermines the effectiveness of semi supervised methods such as pseudo-labeling. In this paper, we propose a Bi-Stream Adversarial Learning network (BiSALnet) to generate pseudo-labels with higher confidence for network optimization. Adversarial learning strategy is adopted in Student and Teacher branches of the two-stream networks. The similarity measurement function minimizes the divergence between the outputs of the Student and Teacher branches, and the discriminator continuously enhances its discriminative ability. In addition, we employ a powerful symmetric positive definite (SPD) manifold structure in the Student branch to capture the desired feature distribution properties. The joint discriminative power of convolutional features and nonlinear complex information aggregated by SPD matrices is combined by the attention feature fusion module to improve the sleep stage classification performance. The BiSALnet is tested on pediatric dataset collected from local hospital. Experimental results show that our method yields the overall classification accuracy of 0.80, kappa of 0.73 and F1-score of 0.76. We also examine the generality of our method on a well-known public dataset Sleep-EDF. Our BiSALnet exhibits noticeable performance with accuracy of 0.91, kappa of 0.85 and F1-score of 0.77. Remarkably, we have obtained comparable performance with state-of-the-art supervised approaches with fairly limited labeled data.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available