Enhancement on Model Interpretability and Sleep Stage Scoring Performance with A Novel Pipeline Based on Deep Neural Network

Considering the natural frequency characteristics in sleep medicine, this paper first proposes a time-frequency framework for the representation learning of the electroencephalogram (EEG) following the definition of the American Academy of Sleep Medicine. To meet the temporal-random and transient nature of the defining characteristics of sleep stages, we further design a context-sensitive flexible pipeline that automatically adapts to the attributes of data itself. That is, the input EEG spectrogram is partitioned into a sequence of patches in the time and frequency axes, and then input to a delicate deep learning network for further representation learning to extract the stage-dependent features, which are used in the classification step finally. The proposed pipeline is validated against a large database, i.e., the Sleep Heart Health Study (SHHS), and the results demonstrate that the competitive performance for the wake, N2, and N3 stages outperforms the state-of-art works, with the F1 scores being 0.93, 0.88, and 0.87, respectively, and the proposed method has a high inter-rater reliability of 0.80 kappa. Importantly, we visualize the stage scoring process of the model decision with the Layer-wise Relevance Propagation (LRP) method, which shows that the proposed pipeline is more sensitive and perceivable in the decision-making process than the baseline pipelines. Therefore, the pipeline together with the LRP method can provide better model interpretability, which is important for clinical support.