Introduction: In critically ill paediatric patients, sleep is essential for recovery and development, yet sleep disturbances are common in the paediatric intensive care unit (PICU), which highlight the need to integrate sleep monitoring into clinical practice. While automated sleep stage classification using machine learning (ML) on single-channel electroencephalography (EEG) data has shown promise in mostly healthy adult populations, its application to critically ill children is challenged by age-specific sleep architecture, medication effects, pathological conditions, and artefacts. This study evaluates whether deep learning (DL) feature extraction and dynamic models can improve sleep staging performance in this population and explores the use of an unsupervised ML model to gain deeper insight into the complex sleep structures in this population.

Methods: This study utilised EEG recordings from three datasets—healthy adults, non-critically ill children, and critically ill children—to train and evaluate supervised and unsupervised sleep stage classification models. As supervised models, a convolutional neural network (CNN) was used for feature extraction, followed by dynamic models including a long short-term memory (LSTM) network and a hidden Markov model (HMM) to account for temporal dependencies in sleep. Additionally, an unsupervised HMM was applied to explore underlying structures in the sleep EEG data without predefined labels.

Results: Supervised models achieved good performance in healthy adults and non-critically ill children, with maximum accuracies of 90.2\% and 77.4\%, respectively, for three-state classification. The added value of dynamic models over the CNN alone varied per dataset and model type and was not consistent. In critically ill children, classification performance was low, with a maximum accuracy of 61.4\%, and notably low macro-F1 and Cohen’s kappa scores (45.9\% and 26.5\%, respectively). The unsupervised HMM revealed that identifying distinct and stable clusters over time was challenging in all datasets. For critically ill children, the model often failed to identify multiple distinct clusters within individual patients, and substantial variability in cluster assignments was observed across patients.

Discussion: This study demonstrates that DL-based feature extraction and dynamic modelling using single-channel EEG can achieve strong sleep staging performance in healthy adults and non-critically ill children, highlighting the potential for (semi-)automated scoring tools in more stable populations. In contrast, performance in critically ill children was notably lower, likely due to factors such as high variability in sleep architecture, signal artefacts, limited data quality, and the uncertain reliability of manually assigned labels. These results suggest that conventional sleep stages do not generalise well to this population, and a purely data-driven, unsupervised approach does not offer a viable alternative. Overall, the findings emphasise the need for a larger dataset of critically ill children, further evaluation of relevant patient and data characteristics, the inclusion of alternative signals such as electrocardiography, and greater focus on model interpretability.

Study objectives: Conventional sleep scoring is based on the scoring criteria of the American Association of Sleep Medicine (AASM) but may not be suited to describe sleep in critically ill children admitted to the Pediatric Intensive Care Unit (PICU). In this study, an anomaly detection model using Gaussian Models trained on sleep stages in data from non-critically ill children is developed to assess if polysomnography(PSG)-derived electroencephalography (EEG) data from critically ill children can be categorized into sleep stages based on these AASM scoring criteria.
Methods: A retrospective study at Erasmus MC Sophia Children’s Hospital, using PSG recordings obtained in non-critically ill children between 2017 and 2021 and in critically ill children between 2020 and 2022.
Gaussian Models were individually trained for each sleep stage using data from non-critically ill children. Anomaly detection was carried out by computing the Mahalanobis Distances and assigning datapoints to specific sleep stages or categorizing them as anomalous. Errors were quantified by calculating the ratio of anomalous epochs to the total number of epochs. The trained Gaussian Models were applied to distinct sleep stages in the data from non-critically ill children. Subsequently, the models were applied to data from critically ill children to determine the categorization of their epochs. This was also analyzed over time and involved comparisons related to medication, mechanical ventilation, and the severity of illness assessed by the PELOD-2 score.
Results: In non-critically ill children the models obtained validation errors aligning with the margin error of the training set. The models could not fully differentiate the distinct sleep stages. In critically ill children, the majority of epochs were classified into multiple sleep stages. High error rates were evident for sleep stages N1, R, and N. Some patients exhibited elevated error rates specifically for sleep stage N1. REM sleep was reduced, consistent with findings from previous studies. In contrast, N3 sleep did not show a reduction. When compared to the sleep stage labels assigned by neurophysiologists, the model classified epochs into multiple sleep stages, while neurophysiologists frequently used the label N. A higher PELOD-2 score did not consistently correlate with an increased occurrence of anomalous classifications in the epochs of these patients to those with lower PELOD-2 scores.
Discussion: Overlap of sleep stages was observed in non-critically ill children. Epochs from critically ill children were classified into multiple sleep stages without clear associations in time or severity of illness. Building upon the established anomaly detection framework is recommended by employing more advanced anomaly detection methods using an informative feature selection. This study marks an initial step, indicating that applying the AASM.

Introduction: Critically ill children admitted to the Paediatric Intensive Care Unit (PICU) have a high risk of disruption of their normal sleep rhythm, which is associated with disturbances in physiology and negative effects on psychological and cognitive functioning. There is a need for real-time, automatic sleep monitoring to minimise disruptions in sleep patterns. The main objective of this thesis was to develop a machine learning model that can classify sleep based on vital signs in critically ill children. In addition, methods were investigated to optimise the decision criteria in multi-class problems.
Methods: Three machine learning algorithms, logistic regression, random forest and extreme gradient boosting (XGBoost), were developed based on both the combination of features extracted from electrocardiogram (ECG) signal and pulse transit time (PTT) and on ECG features alone. To gain insight into the number of sleep stages that could be distinguished, the models were developed for two-class, three-class, four-class and five-class staging. The models were developed, trained and evaluated on a diagnostic dataset (n = 90) containing polysomnography (PSG) measurements of non-critically ill children. During model development, the decision criteria for the different classes were jointly optimised. To evaluate whether the models were generalisable to the PICU population, external validation was performed on a set of 8 of PICU patients.
Results: For each number of sleep stages, the three models performed similarly. However, there was an increase in performance with a decrease in the number of sleep stages with balanced accuracies varying between 0.70 and 0.72 in two-class staging and between 0.41 and 0.42 in five-class staging. External validation on the PICU dataset showed a markedly worse performance for all three models with balanced accuracies varying between 0.56 and 0.62 in two-class staging and between 0.22 and 0.23 in five-class staging.
Conclusion: Machine learning models for sleep classification in children based on vital signs have been developed and show promising results. Nonetheless, the developed models are not generalisable to the PICU population. Further research is recommended with a focus on improving the models such that they can be applied in the PICU. Combining information extracted from vital signs with EEG signal and developing the models directly on PICU data should be considered to improve the performance and thereby contribute to personalised care and minimising sleep disturbances.