Automatic Sleep Stage Scoring on Raw Single-Channel EEG : A comparative analysis of CNN Architectures

Abstract

The significance of sleep in sustaining mental and physiological equilibrium, as well as the relationship between sleep disturbance and disease and death, has long been accepted in medicine. Deep Learning methods have provided State Of The Art performance in tackling numerous challenges in the medical arena since the advent of the domain of HealthTech. Polysomnography is a type of sleep study that uses electroen-cephalogram (EEG) measurements, among other parameters, to get a better picture of a patient’s sleep patterns. Various brain activity correspond to different stages of sleep. Monitoring and interpreting EEG signals and the body’s reactions to the changes in these cycles can help identify disruptions in sleep patterns. Successfully classified sleep patterns can in turn help medical professionals with the prognosis of several pervasive sleep related diseases like sleep apnea and seizures. To address the pitfalls associated with the traditional manual review of EEG signals that help classify sleep stages, in this work, several Convolutional Neural Networks were trained and analysed to classify the five phases od sleep (Wake, N1, N2, N3, N4 and REM by AASM’s standard) using data from raw, single channel EEG signals. With PhysioNet’s Sleep-EDF dataset, this comparative analysis of the performance of popular convolutional neural network architectures can serve as a benchmark to the problem of utilizing EEG data to classify sleep stages. The analysis shows that CNN based methods are adept at extracting and generalizing temporal information, making it suitable for classifying EEG based data.