Automated Infant Cue Classification: a machine learning approach to detecting hunger and feeding discomfort behavioral cues in preterm infants

Abstract

Preterm infants, commonly admitted to hospitals, require intensive and time consuming monitoring. Automated monitoring techniques using video data exist but are not being applied in practice for monitoring infants. By allowing automated monitoring through behavioral cue classification, which infants use to communicate their needs, the burden of monitoring can be relieved allowing for improved health outcomes in preterm infants. This study aims to apply machine learning techniques for automated behavioral cue classification in preterm infants to infer their care needs, specifically of hunger and feeding discomfort. A MoViNet model was trained for this classification problem, selected for their extensive pretraining and multi-class video classification capabilities. Due to the limited availability of labeled data, the techniques of few-shot learning and active learning have been applied to investigate if they improve upon baseline performance. Few-shot learning consists of an initial training phase on similar tasks to allow for quick adjustment of weights. Active learning incorporates additional data labeling, with instances gathered using stratified sampling included in the dataset. It was found that the fully supervised baseline approach was able to successfully uncover patterns in infant behavior. However, few-shot learning resulted in worse performance due to challenges in generalizing from the source to the target domain. Active learning performed comparably to the baseline approach and offered additional value as a labeling tool in the data-scarce setting. The research also revealed the impact of individual differences in behavior, affecting the generalizability of behaviors to other infants and hindering performance. Despite these challenges, individual behavioral differences did not entirely prevent successful classification. By incorporating more training data from new infants, the generalizability of the results and performance could be improved. In sum, this research forms a solid foundation for advancing fully automatic infant monitoring, potentially enabling more individualized care with beneficial health outcomes.