Automatic Sleep Staging based on Curriculum Learning Approach

Abstract

Automatic sleep staging is helpful to improve diagnosis efficiency of sleep-related diseases. This work introduces the many-to-many formulation for automatic sleep staging, which means using a many-to-many mapping to convert the contextual input to the corresponding contextual output. We use convolutional neural networks (CNNs) to perform the many-to-many mapping, and use multilayer perceptron (MLP) to merge the contextual output into the final prediction for a particular epoch. In order to avoid the influence of unobvious characteristic waves and wrong labels on the training process, this work leverages the technology of curriculum learning. By clustering algorithm based on local density, the training set is divided into several subsets according to the signal quality. We design a learning strategy by successively leveraging these subsets. To the best of our current knowledge, this is the first work using curriculum learning for automatic sleep staging. It is showed by experiments that our scheme yields an accuracy comparable to the state-of-the-art on the public dataset Sleep-EDF.