Towards Continuous Sleep Monitoring of Preterm Infants in Complex NICU Scenes: A Video-Based Approach Using Eye Cues

Abstract

A fetus in the third trimester sleeps the majority of the day, which is important for the development of the brain. When born preterm, sleep is likely to be interrupted due to exposure to external stimuli, which could be problematic for development. This study proposes a video-based pipeline to allow for continuous sleep monitoring of preterm infants admitted to the NICU. Such a pipeline would allow for timing of non-essential care, and consequently optimize sleep. Since an infant in the NICU is typically largely occluded, we use only eye cues to classify sleep, as typically at least one eye is visible. The proposed pipeline consists of three main components: an eye localization model, a REM detection model, and a sleep state classification step. The localization model successfully localizes eyes of preterm infants, though it occasionally struggles with closed eyes. Additionally, it is capable of distinguishing between open and closed eyes. The REM model shows promising results, but its performance across train runs varies due to the sensitivity of the training process, likely caused by the limited available data. Ultimately, our results show that sleep state predictions made with our pipeline closely align with ground truth sleep states, indicating potential for practical use. Since eye cues alone are not always reliable for accurately predicting sleep states, we suggest a combination of eye, face, and body cues to ultimately classify sleep more robustly.