Listening Deeper: Neural Networks Unravel Acoustic Features in Preterm Infant Crying

Abstract Infant crying, a noninvasive tool for assessing neurophysiological stress and facilitating communication, has been reported to exhibit atypical acoustic features in preterm infants. However, the majority of previous studies have focused on limited and specific acoustic features, such as fundamental frequency. In this study, for a maximum use of information in infant crying, we employed a convolutional neural network (CNN) approach to gauge whether mel-spectrograms of infants crying capture gestational age variation (79 preterm infants with gestational age [GA] < 37 weeks; 52 term neonates with gestational age ≥ 37 weeks). Our CNN models showed high performances both in binary classifying the pregnancy groups (accuracy = 93.4%) and in estimating the relative and continuous differences in age (r = .73; p < .0001) surpassing the performances in the previous studies. Further inspections of the models revealed that relative differences in gestational age in infant crying were reflected particularly in the temporal features such as prosody. Beyond traditionally assessed acoustic markers, our findings suggest the presence of more complicated features of infant cries tied to neurophysiological states in preterm infants, paving the way for understanding of the early development in preterm infants using deep learning techniques.