Automated Detection of Quiet and Non-Quiet Sleep in Preterm Neonates from aEEG: Towards Predicting Brain Maturation

Abstract

BackgroundPreterm birth is associated with an increased risk for neurodevelopmental impairments, requiring brain monitoring using amplitude-integrated electroencephalography (aEEG). While established tools detect severe dysfunction (e.g., Hellstrom-Westas classification), methods for assessing mild to moderate impairments--such as Burdjalov scoring or expert-based Sleep-Wake Cycle identification--are subjective and require specialized training. Automated neonatal sleep-staging models usually rely on polysomnography from term infants, a resource-intensive method rarely feasible in NICUs, where simplified single-channel aEEG is standard. MethodsaEEG recordings from 40 neurologically healthy neonates (32-42 weeks PMA) were collected and annotated for quiet (QS) and non-quiet sleep (NQS) by an expert clinician. Signals were bandpass filtered, segmented into 30 s epochs, and cleaned using impedance thresholds. 69 temporal, spectral, wavelet, EMG-inspired, and aEEG-envelope features were extracted. The 5 most relevant features were selected for QS/NQS classification using several machine-learning models validated with leave-one-subject-out cross-validation. A partial least squares model was then trained on QS-derived features to predict postmenstrual age and assess correlations with brain maturation. ResultsThe k-Nearest Neighbors (KNN) classifier showed the best QS/NQS discrimination, with mean Cohens{kappa} = 0.69 {+/-} 0.14 for preterm (32-37 weeks PMA) and 0.48 {+/-} 0.21 for term infants. QS-derived features correlated strongly with postmenstrual age (PMA). The PLS model predicted PMA with an average error of 0.88 weeks (MSE = 1.33 weeks, r = 0.91), while the fully automated version using predicted QS segments yielded an error of 1.08 weeks (r = 0.86). ConclusionAutomated QS/NQS detection from single-channel aEEG is feasible in preterm neonates. Despite reduced accuracy in term infants, QS-derived features closely track brain maturation, supporting the potential of aEEG-based models for objective, early detection of neuromaturation delays in preterm infants