Automatic Classification of Sleep/Wake States of Newborns Using Deep Learning

Abstract

In the neonatal intensive care unit (NICU), various medical procedures are performed day and night in an incubator, and since the NICU environment is very different from that in the womb, it is necessary to investigate the effects on the neurodevelopment of newborn infants. Currently, the sleep-wake state is evaluated visually by nurses and with an actigraph, a device attached to the leg of the newborn. In this study, we propose a method for automatically estimating the sleep-wake state of newborns using video images based on Brazelton’s classification. Experiments are conducted using deep learning 3DCNN with four methods: inputting training data as it is, shifting frames extracted from the input data to double the number of training data, increasing the number of training data by a factor of six, and generating difference images. The experimental results show that the highest macro-F1 value (0.766) was obtained when the dataset was trained as is, without data expansion or difference images. The results show that the accuracy of each method can be improved by using deep learning.