Aims: Sleep monitoring is essential for assessing neurophysiological maturation in premature newborns within Neonatal Intensive Care Unit (NICU), highlighting the need for automatic sleep state classification. Current approaches rely on either single-modality signals (ECG, EEG) or multimodal features with time-independent models. This study proposes a temporal deep learning approach to estimate Quiet Sleep (QS) using cardiorespiratory and motion-based features, as QS is a key neurological marker.
Methods: Data of 47 preterm newborns with annotated recordings over six days were included (2584 hours in total). From 120 initial features extracted from cardiorespiratory signals (in time, frequency and non-linear domains) and video-based motion analysis, hierarchical clustering was applied to group correlated features, resulting in 32 selected features, supplemented by gestational and postmenstrual age. Sequential 40-second data windows were processed by a Long Short-Term Memory (LSTM) network with a multi-head temporal attention mechanism. Performance was evaluated via 5-fold cross-validation, with folds generated using a simulated annealing algorithm to ensure balanced subject and recording distributions. Following the only prior work applying an LSTM to cardiac data for sleep state classification, an ECG-only experiment was also performed.
Results: The ECG-only model achieved a balanced accuracy of 0.76 ± 0.01 and a Cohen's kappa of 0.52 ± 0.01, while the multimodal LSTM improved performance to 0.80 ± 0.01 and 0.59 ± 0.01, respectively. Permutation importance analysis indicated that the most influential features were the number of motion intervals, the minimum duration, and the coefficient of variation of RR intervals, as well as the coefficient of variation of BB intervals in respiration.
Conclusion: This study demonstrates that jointly modeling the temporal dynamics of cardiorespiratory and motion features via an LSTM network with attention mechanism significantly improves automatic QS detection in preterm newborns. This framework offers a promising and non-invasive tool for continuous sleep monitoring in NICU settings.