U-Time: A Fully Convolutional Network for Time Series Segmentation\n Applied to Sleep Staging

Abstract

Neural networks are becoming more and more popular for the analysis of\nphysiological time-series. The most successful deep learning systems in this\ndomain combine convolutional and recurrent layers to extract useful features to\nmodel temporal relations. Unfortunately, these recurrent models are difficult\nto tune and optimize. In our experience, they often require task-specific\nmodifications, which makes them challenging to use for non-experts. We propose\nU-Time, a fully feed-forward deep learning approach to physiological time\nseries segmentation developed for the analysis of sleep data. U-Time is a\ntemporal fully convolutional network based on the U-Net architecture that was\noriginally proposed for image segmentation. U-Time maps sequential inputs of\narbitrary length to sequences of class labels on a freely chosen temporal\nscale. This is done by implicitly classifying every individual time-point of\nthe input signal and aggregating these classifications over fixed intervals to\nform the final predictions. We evaluated U-Time for sleep stage classification\non a large collection of sleep electroencephalography (EEG) datasets. In all\ncases, we found that U-Time reaches or outperforms current state-of-the-art\ndeep learning models while being much more robust in the training process and\nwithout requiring architecture or hyperparameter adaptation across tasks.\n