Listening deeper: neural networks unravel acoustic features in preterm infant crying

Abstract

Abstract Early infant crying provides critical insights into neurodevelopment, with atypical acoustic features linked to conditions such as preterm birth. However, previous studies have focused on limited and specific acoustic features, hindering a more comprehensive understanding of crying. To address this, we employed a convolutional neural network to assess whether whole Mel-spectrograms of infant crying capture gestational age (GA) variations (79 preterm infants; 52 term neonates). Our convolutional neural network models showed high accuracy in classifying gestational groups (92.4%) and in estimating the relative and continuous differences in GA ( r  = 0.73; p  < 0.0001), outperforming previous studies. Grad-CAM and spectrogram manipulations further revealed that GA variations in infant crying were prominently reflected in temporal structures, particularly at the onset and offset regions of vocalizations. These findings suggest that decoding spectrotemporal features in infant crying through deep learning may offer valuable insights into atypical neurodevelopment in preterm infants, with potential to enhance early detection and intervention strategies in clinical practice.