Automatic Classification of Neonatal Sleep-Wake States Based on Facial Video Analysis

Abstract

Premature babies are admitted to the NICU (Neonatal Intensive Care Unit) for several weeks and generally placed under high medical supervision. To provide a better environment to them, some researchers investigate the affection of light and noise in the NICU on the formation of the sleep-wake cycle of the newborn called Circadian rhythm. These researches require the optimal evaluation method of the sleep-wake state. The visual assessment by nurses do not guarantee enough inter-tester reliability, and the measurement puts an additional burden on them. The conventional sleep-wake states discrimination method requires attachment devices on the subject’s body. This paper proposes the automatic classification method of the sleep-wake states of neonates by using only facial information. In this research, we extract gradient features and spatio-temporal HOGV features from 3,600 face image frames (1 minute). According to Blazelton’s method, this study classifies the sleep-wake states into six classes by using machine learning techniques. Support Vector Machine and Random Forest were used in the experiment. The spatio-temporal HOGV feature is an extension of the HOG feature to the time domain. The experiments using two kinds of feature quantities and classifiers showed that the highest accuracy rate (54.4%) was obtained by the gradient feature and Random Forest. This result suggested the possibility of improving accuracy by combining facial information with body movement and other conventional features.