Abstract
Sleep is associated with a variety of chronic diseases as well as most psychiatric, addiction and mood disorders. To analyze sleep patterns in rodents, researchers analyze polysomnogram data containing electroencephalographs (EEG) and electromyographs (EMG). However, the analysis is performed manually by a expert human scorer, which is a slow, time consuming, and expensive process that is also subject to known human error and inter-scorer inconsistency [1]. To address this, researchers have developed a variety of techniques to automatically classify rodent sleep states using features extracted from EEG and EMG signals [2]. In many approaches, researchers extract a variety of heuristic features from explicitly chosen spectral bands of the EEG and EMG signals [3]. However, human designed, heuristic features often do not capture complete salient sleep-state information, which leads to inferior classification performance.