Automatic estimation of sleep states in preterm neonates by multimodal information fusion

Overview & Synopsis

Sleep plays an important role in the neurological development of newborns. In preterm infants, its quality and organization are early indicators of brain maturation and potential developmental disorders. In Neonatal Intensive Care Units (NICUs), assessing sleep states is essential, but traditional methods based on observation or polysomnography remain cumbersome and poorly suited to continuous monitoring. Many studies have proposed automatic methods using physiological or behavioral signals, yet their robustness, interpretability, and generalizability remain limited. This thesis develops a complete pipeline for sleep state classification, focusing on Quiet Sleep (QS), a key stage of neurological development. The approach includes: i) feature extraction from three modalities (ECG, respiration, movements); ii) an unsupervised classification to support QS annotation; iii) a compact and interpretable supervised model to estimate QS and distinguish three states (QS, non-quiet sleep, wake). Results show good agreement with expert annotations for QS, while three-state classification still requires improvement. This approach paves the way for non-invasive, automatic monitoring of preterm newborns’ sleep, especially QS.