Deep learning in classifying sleep stages

Abstract

This paper presents a deep feed-forward neural network classifier to automatically classify the stages of sleep using raw data taken from a single electropalatogram channel (Fpz-Cz). No features are extracted at all from the data, and the network can classify the five sleep stages: waking, Nl, N2, N3, N4, and rapid eye movement. The network has three layers, takes as an input a l-s epochs to be classified, and requires no signal pre-processing nor feature extraction. We trained and evaluated our system using DeepLearning4J, the free Java framework for test data taken from PhysioNet's Polysomnography Sleep database. An accuracy of 0.99 within a constrained environment has been reached.