[home] [Personal Program] [Help]

---

Session: Birth & Neonates
Session starts: Thursday 24 January, 13:30
Presentation starts: 13:30
Room: Lecture room 558

---

Ofelie De Wel (Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics, KU Leuven, Leuven, Belgium)
Mario Lavanga (Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics, KU Leuven, Leuven, Belgium)
Alexander Caicedo (Department of Applied Mathematics and Computer Science, Faculty of Natural Sciences and Mathematics, Universidad del Rosario, Bogotá, Colombia)
Katrien Jansen (Department of Development and Regeneration, University Hospitals Leuven, Neonatal Intensive Care Unit & Child Neurology, KU Leuven, Leuven, Belgium)
Gunnar Naulaers (Department of Development and Regeneration, University Hospitals Leuven, Neonatal Intensive Care Unit, KU Leuven, Leuven, Belgium)
Sabine Van Huffel (Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics, KU Leuven, Leuven, Belgium)

Abstract:
Preterm infants are at increased risk of neurodevelopmental impairment. Therefore, these vulnerable neonates need close monitoring of their neurological function during the first critical weeks. Sleep state organization undergoes fast development in this period and reflects the level of functional brain integrity. Hence, an automated identification of neonatal sleep stages is of great interest to evaluate electrocortical maturation. This study aimed to discriminate quiet sleep (QS) from non-quiet sleep (non-QS) using the nonlinear dynamics of the EEG signal. The analysis is performed on 97 multichannel EEGs from 26 prematurely born babies recorded at a postmenstrual age between 27 and 42 weeks. Multiscale entropy, which assesses the degree of regularity of a signal at multiple time scales, is used to quantify the complexity of each 8-channel EEG segment. The entropy values of each EEG recording are then organised in a third order tensor with modes: channels, scales and time segments. Subsequently, a rank-1 nonnegative canonical polyadic decomposition of the multiscale entropy tensor is computed, resulting in a spatial signature, a scale signature and a temporal signature. The temporal signature captures the main variation of the EEG complexity over the different time segments. After smoothing this factor with a moving average filter, k-means clustering is applied to obtain two distinct clusters. The silhouette coefficient is computed as an internal evaluation criterion for the quality of the clustering, and has an average value of 0.81. In order to confirm that the two clusters correspond to QS and non-QS periods, external evaluation of the clustering using the visual sleep stage scoring by 2 clinicians is performed as well. Comparison with the clinical labels revealed that the neural complexity is significantly lower in QS compared to non-QS. This prior information is used to label the clusters and the average sensitivity, specificity, accuracy and area under the curve are equal to 72%, 80%, 78% and 84%, respectively. This study shows that the complexity of brain dynamics exhibits fundamental differences between vigilance states in preterm newborns. Moreover, a novel unsupervised approach to detect quiet sleep based on the tensor factorization of the multiscale entropy tensor is proposed.