Automated sleep-wake cycling detection in neonates from cerebral function monitor signals.
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis presents a link between a clinical need involving the analysis of high frequency physiological data and the informatics and technology designed to automate these specific clinical rules. An algorithmic design for the automated detection of sleep-wake cycling patterns in neonates aged 29-44 weeks gestational age via cerebral function monitor signal is presented in this thesis. The design also includes the automation of impedance levels and background classification of electrical activity via cerebral function monitor signal. The relevance of these algorithms is demonstrated within the neonatal intensive care unit as this monitor is commonly used at the bedside of critically ill infants. The design composition determines analyzable and clinically relevant data through the assessment of impedance levels associated with the cerebral function monitor signal, then the overall background classification of the infant’s cerebral electrical activity which indicates whether or not sleep-wake cycling can be present in the signal trace. A third output is the detection of sleep-wake cycling. Current practice in an intensive care setting involves the meticulous and time-consuming process of manual interpretation by a health care professional, subsequently the automation of these processes has the potential to reallocate the use of resources in the form of staff, increase the rate at which the current practice takes place, improve the timing of medical intervention to allow for maximal neurological development in patients as well as facilitate comprehensive analysis with other physiological data streams in unison to deliver enhanced decision making. Performance of the algorithms in comparison to expert clinical annotation resulted in concordance values of 95.70% for impedance, 78.49%, 81.25% sensitivity and 75.32% specificity for sleep-wake cycling and finally 76.34% for background classification. Through this retrospective analysis of de-identified patient data it was determined that this can be applicable to a real-time computer software environment enabled by stream computing.