To develop a generalized automated sleep staging method based on the gold standard modality, electroencephalograms (EEGs), requires a large and accurately labeled training and test set acquired from different individuals with diverse demographics and medical conditions. However, data in the training set may exhibit changes in the EEG patterns that are very different from the data in the test set, due to inherent inter-subject variability, electrode misplacement, and the variability of medication use/response. Training an algorithm on such data without accounting for this diversity can lead to underperformance and a lack of generalizability on novel data. Previous methods have attempted to address this by developing robust representations across all individuals in the dataset using deep transfer learning approaches. However, not all parts of the training data are as relevant as others to the test data. Forcing the alignment of these nontransferable data with the transferable data may lead to a negative impact on the overall performance. This work jointly learns patient-invariant representations and weights features (spectrogram coefficients) to enhance the contribution of relevant features in the final model and decrease the impact of irrelevant features using an unsupervised approach. The proposed method leverages transferable and discriminable knowledge from the training set to the test set.
Using a large public database of 42,560 hours of EEG, recorded from 5,793 from Sleep Heart Health Study, we demonstrate that adversarially learning a network with an importance weighting scheme significantly boosts performance compared to state-of-the-art deep learning approaches in the cross-subject scenario. The proposed method improves, on average, accuracy from 0.81 to 0.94, precision from 0.81 to 0.82, and sensitivity from 0.74 to 0.85.