Neonatal sleep state identification using deep learning autoencoders

Abstract

Neonatal sleep state analysis provides a tool for diagnosis of several possible physiological disorders in newborns. The sleep state identification is a time consuming procedure where the sleep state is determined in epochs of 60 second for an entire sleep recording. A new technique for automated sleep state identification in neonates is proposed. The proposed system comprises two major step; feature extraction and classification. Twelve features were extracted from a single EEG recording. The features extracted were based on statistical parameters extracted from both temporal and spectral domains of the EEG signal. The total number of recordings used was 29 EEG recordings acquired from newborns (14 preterm infants and 15 fullterm). The classification was done based on deep autoencoder neural networks. The structure of the network used was two autoencoder layers and one softnet output layer. The performance of the proposed system was evaluated for both preterm and fullterm records assembled in one group of data using 10 fold cross validation. Also, the performance was tested for the two groups separately. The reported accuracy was 0.804 for the entire data sets. Keywords: Sleep, Feature extraction, Pediatrics, Electroencephalography, Machine learning, Neural networks, Signal processing