Capturing the Development of Internal Representations in a High-Performing Deep Network for Sleep Stage Classification
AbstractDeep neural networks can provide accurate automated classification of human sleep signals into sleep stages that enables more effective diagnosis and treatment of sleep disorders. We develop a deep convolutional neural network (CNN) that attains state-of-the-art sleep stage classification performance on input data consisting of human sleep EEG and EOG signals. Nested cross-validation is used for optimal model selection and reliable estimation of out-of-sample classification performance. The resulting network attains a classification accuracy of $$84.50 \pm 0.13\%$$ 84.50 ± 0.13 % ; its performance exceeds human expert inter-scorer agreement, even on single-channel EEG input data, therefore providing more objective and consistent labeling than human experts demonstrate as a group. We focus on analyzing the learned internal data representations of our network, with the aim of understanding the development of class differentiation ability across the layers of processing units, as a function of layer depth. We approach this problem visually, using t-Stochastic Neighbor Embedding (t-SNE), and propose a pooling variant of Centered Kernel Alignment (CKA) that provides an objective quantitative measure of the development of sleep stage specialization and differentiation with layer depth. The results reveal a monotonic progression of both of these sleep stage modeling abilities as layer depth increases.