Intelligent monitoring and interpretation of preterm physiological signals using machine learning

Thumbnail Image
thesis_final_Jan02.pdf(8 MB)
Full Text E-thesis
Semenova, Oksana
Journal Title
Journal ISSN
Volume Title
Published Version
Research Projects
Organizational Units
Journal Issue
Every year, more than one in ten babies are born prematurely. In Ireland of the 70000 babies delivered every year, 4500 are born too early. Premature babies are at a higher risk of complications, which may lead to both short-term and long-term adverse health outcomes. The neonatal population is especially vulnerable and a delay in the identification of medical conditions, as well as delays in the initiating the correct treatment, may be fatal. After birth, preterms are admitted to the neonatal intensive care unit (NICU), where a continuous flow of information in the form of physiological signals is available. Physiological signals can assist clinicians in decision making related to the diagnosis and treatment of various diseases. This information, however, can be highly complex, and usually requires expert analysis which may not be available at all times. The work conducted in this thesis develops a decision support systems for the intelligent monitoring of preterms in the NICU. This will allow for an accurate estimation of the current health status of the preterm neonate as well as the prediction of possible long-term complications. This thesis is comprised of three main work packages (WP), each addressing health complication of preterm on three different stages of life. At the first 12 hours of life the health status is quantified using the clinical risk index for babies (CRIB). This is followed by the assessment of the preterm’s well-being at discharge from the NICU using the clinical course score (CCS). Finally, the long-term neurodevelopmental follow-up is assessed using the Bayley III scales of development at two years. This is schematically represented in Figure 1 along with the main findings and contributions. Low blood pressure (BP) or hypotension is a recognised problem in preterm infants particularly during the first 72 hours of life. Hypotension may cause decreased cerebral perfusion, resulting in deprived oxygen delivery to the brain. Deciding when and whether to treat hypotension relies on our understanding of the relation between BP, oxygenation and brain activity. The electroencephalogram (EEG) is the most commonly used technology to assess the ‘brain health’ of a newborn. The first WP investigates the relationship between short-term dynamics in BP and EEG energy in the preterm on a large dataset of continuous multi-channel unedited EEG recordings in the context of the health status measured by the CRIB score. The obtained results indicate that a higher risk of mortality for the preterm is associated with a lower level of nonlinear interaction between EEG and BP. The level of coupling between these two systems can potentially serve as an additional source of information when deciding whether or not to intervene in the preterm. The electrocardiogram (ECG) is also routinely recorded in preterm infants. Analysis of heart rate variability (HRV) provides a non-invasive assessment of both the sympathetic and parasympathetic control of the heart rate. A novel automated objective decision support tool for the prediction of the short-term outcome (CCS) in preterm neonates who may have low BP is proposed in the second WP. Combining multiple HRV features extracted during hypotensive episodes, the classifier achieved an AUC of 0.97 for the task of short-term outcome prediction, using a leave-one-patient-out performance assessment. The developed system is based on the boosted decision tree classifier and allows for the continuous monitoring of the preterm. The proposed system is validated on a large clinically collected dataset of multimodal recordings from preterm neonates. If the correct treatment is initiated promptly after diagnosis, it can potentially improve the neurodevelopmental outcome of the preterm infant. The third WP presents a pilot study investigating the predictive capability of the early EEG recorded at discharge from the NICU with respect to the 2-year neurodevelopmental outcome using machine learning techniques. Two methods are used: 1) classical feature-based classifier, and 2) end-to-end deep learning. This is a fundamental study in this area, especially in the context of applying end-to-end learning to the preterm EEG for the problem of long-term outcome prediction. It is shown that for the available labelled dataset of 37 preterm neonates, the classical feature-based approach outperformed the end-to-end deep learning technique. A discussion of the obtained result as well as a section highlighting the possible limitations and areas that need to be investigated in the future are provided.
Hypotension , Outcome prediction , Preterm neonate , Machine learning , Mutual information , Boosted decision trees
Semenova, O. 2020. Intelligent monitoring and interpretation of preterm physiological signals using machine learning. PhD Thesis, University College Cork.