Attention on Sleep Stage Specific Characteristics

Manual sleep stage classification relies on visual inspection of 30-second windows comprising multi-sensor measurements The ability of neural networks to model complex relations has made them a popular, faster, alternative. However, it often remains unclear which parts of the data predominantly contributed to the model's decision. This is especially ambiguous in sleep staging, where the coarse labeling per 30-second windows may assign mixtures of class-specific features to a single class. To boost the transparency of deep neural classifiers, we propose a dynamic discrete attention module that actively selects the subset of the input space aligned with the class label. The module can be combined with a typical classification network, and may additionally serve as a data-driven tool to discover sleep stage specific features in polysomnography data. We validate the method on synthetic and patient data. We observe that only a small subset of data from the 30-second window is required to retain accurate classification, and that the attention mechanism boosts performance. Analysis of the dynamic attention masks, moreover, shows clear sleep stage adaptive channel selection.