Automatic Classification of Sleep-Wake States of Newborns Using Only Body and Face Videos.

The premature newborn receives specialized medical care in the neonatal intensive care unit (NICU), where various medical devices emit excessive light and sound stimulation, and those prolonged exposures to stimuli may cause stress and hinder the development of the newborn's nervous system. The formation of their biological clock or circadian rhythm, influenced by light and sound, is crucial for establishing sleep patterns. Therefore, it is essential to investigate how the NICU environment affects a newborn's sleep quality and rhythms. Brazelton's classification criteria measure the sleep-wake state of newborns, but the visual classification is time-consuming. Therefore, we propose a method to reduce the burden by automatically classifying the sleep-wake state of newborns from video images. We focused on videos of whole-body and face-only videos of newborns and classified them into five states according to Brazelton's classification criteria. In this paper, we propose and compare methods of classifying whole-body and face-only videos separately using a three-dimensional convolutional neural network (3D CNN) and combining the two results obtained from whole-body and face-only videos with time-series smoothing. Experiments using 16 videos of 8 newborn subjects showed that the highest accuracy of 0.611 and kappa score of 0.623 were achieved by weighting the time-series smoothed results from whole-body and face-only videos by the output probabilities from the 3D CNN. This result indicated that the time-series smoothing and combining the results based on probabilities is effective. [ABSTRACT FROM AUTHOR]