Current approaches to developing a generalized automated sleep staging method rely on constructing a large labeled training and test corpora by leveraging electroencephalograms (EEGs) from different individuals. However, data in the training set may exhibit changes in the EEG pattern that are very different from the data in the test set due to inherent inter-subject variability, heterogeneity of acquisition hardware, different montage choices and different recording environments. Training an algorithm on such data without accounting for this diversity can lead to underperformance. In order to solve this issue, different methods are investigated for learning an invariant representation across all individuals in datasets. However, all parts of the corpora are not equally transferable. Therefore, forcefully aligning the nontransferable data may lead to a negative impact on the overall performance. Inspired by how clinicians manually label sleep stages, this paper proposes a method based on adversarial training along with attention mechanisms to extract transferable information across individuals from different datasets and pay attention to more important or relevant channels and transferable parts of data, simultaneously. Using two large public EEG databases - 994 patient EEGs (6,561 hours of data) from the Physionet 2018 Challenge (P18C) database and 5,793 patients (42,560 hours) EEGs from Sleep Heart Health Study (SHHS) - we demonstrate that adversarially learning a network with attention mechanism, significantly boosts performance compared to state-of-the-art deep learning approaches in the cross-dataset scenario. By considering the SHHS as the training set, the proposed method improves, on average, precision from 0.72 to 0.84, sensitivity from 0.74 to 0.85, and Cohen’s Kappa coefficient from 0.64 to 0.80 for the P18C database.
展开▼
