The classification of neonatal sleep states using NIRS: what features are valuable?

Zahn, Maren

The classification of neonatal sleep states using NIRS: what features are valuable?

Files

Zahn, M. s-4825489-MSc-MKI92-Thesis-2023.pdf (3.83 MB)

Authors

Zahn, Maren

Issue Date

2023-06-01

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16430

Abstract

Proper sleep is of great importance for preterm infants, as sleep is linked to brain development. Thus, a wide range of methods to automatically assess preterm sleep have been developed. However, most of those methods employ signals that are recorded by attaching adhesive electrodes to the infant’s skin, which can damage the skin of the vulnerable preterm infants. An alternative approach involves the use of a near-infrared spectroscopy (NIRS) sensor, which can be attached without harming the infant’s skin. The objective of this study is to investigate the potential of NIRS derived features to automatically classify preterm sleep. To do so, the following features are extracted from the NIRS signal: oxygenated hemoglobin (HbO2) concentrations, deoxygenated hemoglobin (HHb) concentrations, tissue saturation index, the negative correlation between HbO2 and HHb, motion related features, heart rate and respiratory rate. These features are the input to a random forest and a support vector machine (SVM) for the final classification. To train and test the random forest and SVM, this study used NIRS data recorded from 9 subjects admitted to the neonatal intensive care unit of the Wilhelmina’s children’s hospital. The SVM demonstrates the best classification results, successfully distinguishing active and quiet sleep with a sensitivity of 0.72 and a specificity of 0.86. Furthermore, various subsets of features have been tested to evaluate feature importance. A combination of heart rate, respiratory rate and motion related features yielded the best results from all evaluated combinations. Thus, heart rate, respiratory rate and motion related features are the most valuable features for the classification task. The obtained promising results indicate that NIRS has the potential to be used to assess preterm sleep.