Improving Robustness and Reliability in Neonatal Brain Monitoring

Approximately 4.3% of newborns in Belgium require care in the neonatal intensive care unit (NICU) due to various reasons, with preterm birth being the most prevalent cause. In these neonates, monitoring the brain is crucial as brain injury is common. Two non-invasive brain monitoring techniques commonly used in the NICU are electroencephalography (EEG) and near-infrared spectroscopy (NIRS). EEG measures brain activity, while NIRS measures brain oxygenation. Analysing these signals aids in detecting brain injuries, guiding clinical interventions, and optimizing brain perfusion. However, visual analysis of these signals is time-consuming, requires expertise, and can be subjective. Automated analysis using advanced signal processing and machine learning techniques can provide objective and fast analysis of EEG and NIRS signals, both in real-time and off-line settings. Therefore, such automated methods hold significant potential as valuable tools in the NICU.
Applications of such automated analysis covered in this thesis are the assessment of neurovascular coupling, estimation of brain maturation, and detection and analysis of sleep. Recent advances have led to automated methods for these applications, but challenges regarding robustness and reliability are currently impeding their implementation in clinical practice.

This thesis aims to address these challenges and focusses on developing robust algorithms for neonatal EEG and NIRS, which can be used for reliable decision-making support in the NICU. The work in this thesis can be categorized into three parts, each covering one of the following objectives: 
1) the development of advanced signal processing methods for the analysis of neonatal EEG and NIRS monitoring,
2) the development of robust algorithms for reliable analysis of long neonatal EEG recordings, and
3) the assessment of brain maturation and sleep in neonates with normal and abnormal outcomes.

The first part of this thesis focusses on developing advanced signal processing methods for the analysis of neonatal EEG and NIRS monitoring. First, a wavelet-based method for quantifying EEG-NIRS coupling is described.  This method aims to quantify the interplay between cerebral activity and perfusion, which can be interpreted as a measure of neurovascular coupling (NVC). In order to improve the robustness of this method, we proposed extensions that improve the handling of artefacts and missing data, and correct for confounding changes in arterial oxygen saturation. The enhanced methodology was evaluated on a small dataset of neonates undergoing therapeutic hypothermia, and the amount of detected NVC was related to brain injury.  A combination of EEG and NVC biomarkers has the potential to predict brain injury accurately, but the results should be validated in larger datasets. Besides this method for quantifying EEG-NIRS coupling, we investigated an advanced method for analysing multi-channel EEG that is emerging in the neonatal field: microstate analysis.  By investigating microstates in a dataset of preterm neonates, we showed that this data-driven analysis framework was capable of capturing changing global brain dynamics in the developing neonatal EEG.  These results provide a reference for future research and demonstrate the potential of data-driven techniques for analysing patterns in multi-channel EEG data.

The second part of the thesis is dedicated to improving the robustness of automated EEG-based methods for brain maturation and sleep analysis. As a first step, we developed an automated method for the detection of artefacts in neonatal EEG, which are a common source of error when doing automated analysis of long, continuous EEG recordings. A semi-supervised deep learning approach was proposed to learn from both labelled and unlabelled data, which improved performance, especially when the amount of labelled data was limited. In the remainder of this second part of the thesis, the added value of automated artefact detection is demonstrated by integrating it into algorithmic pipelines designed for assessing brain maturation and detecting sleep patterns. Here, we also incorporated methods for novelty detection and uncertainty quantification which further improved the robustness and reliability of the automated EEG analyses.
We showed that the robust algorithms may not have a large effect on average performance, but do result in more reliable results on an individual level.

In the third part, we evaluate the clinical use of the previously mentioned robust algorithms for brain maturation and sleep analysis. First, we applied the automated EEG analyses to a group of neonates with congenital heart diseases. 
The analysis indicated significant brain maturation delays in these patients, as well as changes in sleep patterns. The algorithms were also applied to a larger dataset comprising (extreme) preterm and healthy term neonates. Here, significant deviations in the EEGs of the extreme preterm group were observed, indicating delayed or atypical brain maturation. Our results showed that some EEG-derived features were modestly related to long-term neurodevelopmental outcome, but further investigation is needed to confirm our findings. Overall, the studies presented in this third part of the thesis demonstrate the potential of automated EEG analysis as a valuable tool for studying EEG and identifying abnormalities in brain function and development.

In summary, various advanced tools exist for the automated analysis of neonatal EEG and NIRS. However, before these algorithms can be used in a fully-automated way, the robustness and reliability in a clinical setting need to be guaranteed. In this thesis, we made several contributions that gave rise to robust algorithms for reliable automated analysis of EEG and NIRS. The sensitivity of these methods to abnormal brain function and development suggests that these algorithms may play a future role in computer-aided diagnoses and prognoses in the NICU.