Development of ECG-based machine learning methods for early detection of neonatal brain injury
Loading...
Files
Full Text E-thesis
Date
2026-01-04
Authors
Rezaei, Kimia
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
Hypoxic-Ischemic Encephalopathy (HIE) is caused by inadequate blood flow and oxygen delivery to the brain, typically occurring around the time of birth and can lead to neurological impairments or death. HIE is the most common underlying cause of neonatal seizures, many of which are subclinical and difficult to detect. Early and accurate assessment of both HIE severity and seizure activity is essential in the neonatal intensive care unit (NICU) to support immediate diagnosis and treatment.
The increasing use of the Electroencephalogram (EEG) for supporting clinical decisions, such as initiating therapeutic hypothermia, seizure detection and prognostication, led to growing interest in developing automated systems to assist clinicians in grading HIE. Monitoring EEG requires specialist expertise and is resource-intensive, limiting timely diagnosis in many low-resource settings. The Electrocardiogram (ECG), however, is routinely monitored and is widely accessible.
This thesis investigates various approaches for the two tasks of HIE grading and seizure detection, utilizing a large ECG dataset from newborns. A traditional approach is implemented by deriving the heart rate (HR) from the ECG. The most effective HR variability (HRV) features are analysed, and the most powerful classification models are benchmarked.
In contrast, two new approaches are proposed. By eliminating the ECG preprocessing step and leveraging more robust deep learning models, computational overhead can be reduced, enabling real-time processing, and enhancing the practicality of HIE grading and seizure detection for clinical use. The first method introduced spectrogram calculated directly from the raw ECG data. Lightweight convolutional neural network (CNN) architectures are designed, tailored to the input spectrogram to effectively capture hierarchical representations through progressively expanding receptive fields.
The second method proposes trend-augmented multiscale tokenization transformer (TMFormer), a novel transformer-based model that introduces a tokenization approach derived from raw ECG signal attributes, capturing spatial patterns at multiple scales and tracking the signal's long-term progression. Since seizure detection likely depends on short ECG segments, where derived representations such as spectrograms capture more informative patterns than raw signals, we developed a vision transformer with convolutional patch embedding (CNN–ViT). HIE grading results represent a promising step toward developing HIE assessment tools to support clinicians in timely decision-making for infant treatments even in resource-limited hospitals, However, Seizure detection remains challenging.
Description
Keywords
Hypoxic ischaemic encephalopathy , Electroencephalogram , Electrocardiogram , Neonatal , Machine learning , Spectrogram , Hear rate features , Transformer model , Convolutional neural network
Citation
Rezaei, K. 2026. Development of ECG-based machine learning methods for early detection of neonatal brain injury. PhD Thesis, University College Cork.