Neonatal brain injury detection using advanced deep learning architectures
Loading...
Files
Full text E-thesis
Date
2026-03-10
Authors
Yu, Shuwen
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
Hypoxic-ischemic encephalopathy (HIE) is a severe neonatal brain injury resulting from a lack of oxygen (hypoxia) and inadequate blood flow (ischemia) to the brain around the time of birth. It remains one of the leading causes of neonatal mortality and long-term neurodevelopmental morbidity worldwide. Early, accurate and continuous assessment is critical for guiding neuroprotective interventions. However, conventional clinical observation is often subjective and limited by the availability of expert interpretation. This highlights the urgent need for a paradigm shift toward data-driven, multimodal physiological monitoring and intelligent decision-support systems. Artificial intelligence (AI) assisted models offer the potential to support early diagnosis, guide timely intervention, and enable continuous 24/7 monitoring when clinical expertise is not immediately available.
This thesis presents deep learning based algorithms for neonatal HIE classification using electroencephalography (EEG) and heart rate variability (HRV) signals. With the objective of providing accurate predictions from different modalities, multiple of training paradigms, including fully supervised, weakly supervised and self-supervised learning, were explored. The developed models progressively evolve towards a unified architectural framework, compatible with different physiological modalities. In addition, large scale EEG and HRV datasets were constructed to support the development and evaluation of data-intensive deep learning models.
First, a fully supervised learning system based on a lightweight Fully Convolutional Network (FCN) architecture was proposed for multichannel EEG-based neonatal HIE severity classification. The model directly processes raw EEG signals without manually engineered features and maintains a simpler structure with fewer parameters compared to the conventional Convolutional Neural Network (CNN) baselines. Trained on the expert-annotated ANSeR2 EEG dataset (315 one-hour epochs), the model was evaluated on an independent strongly labelled ANSeR1 dataset (338 one-hour epochs) under cross-dataset and mismatched conditions. Despite this challenge, the proposed method achieved 86.3\% AUC and 86.9\% accuracy on the 4-class test set, representing a 23\% relative improvement in test accuracy over the best-performing CNN baseline.
Second, to construct a high quality HRV dataset, an enhanced Pan-Tompkins algorithm was developed to improve R-peak detection robustness in noisy EEG recordings. This enhancement significantly increased the availability and quality of extracted heart rate signals. A hybrid Convolution-Transformer architecture, termed HRVConformer, was introduced for HRV-based HIE classification. By combining convolutional \redmk{layers} for local feature extraction with Transformer-based attention for global context modelling, the architecture effectively enhanced the signal representation and classification performance. This model was trained on a large ANSeR2 HRV dataset consisting of 1,573 one-hour epochs, (including 257\,h of expert-annotated data and extensive weakly labelled data) and evaluated on another independent ANSeR1 HRV test set including 215 one-hour expert-annotated epochs. The proposed approach achieved an AUC of 83.23\% and accuracy of 74.56\% on the 2-class classification problem -- is therapeutic hypothermia recommended? This outperformed the Transformer, ResNet50, and fully convolutional network baselines.
Finally, to address the limited availability of expert annotations, a unified self-supervised Mask Autoencoder framework with Conformer backbone, termed MAEConformer, was proposed for large-scale pretraining on unlabelled EEG and HRV data. By reconstructing masked segments of input signals, the model learned intrinsic physiological representations in a task-agnostic manner. A joint optimization strategy combining time-domain reconstruction loss and multi-resolution Short-time Fourier Transform based frequency-domain loss was introduced to enhance the quality of the spectral representations. The models were pretrained on 6,030 hours unlabelled multichannel EEG and 4,868 hours of unlabelled HRV datasets, respectively. After fine-tuning and linear probing on smaller expert-annotated datasets, the MAE-EEG model achieved test AUCs of 96.56\% (four-class) and 97.17\% (two-class) on the ANSeR1 EEG test set at the one-hour epoch-level, surpassing a range of supervised and self-supervised baselines and achieving state-of-the-art performance. For HRV-based classification, the MAE-HRV achieved 82.42\% test AUC on the two-class ANSeR1 HRV test set, slightly below the HRVConformer baseline but demonstrating strong training stability and data efficiency.
Description
Keywords
Hypoxic-ischemic encephalopathy , Electroencephalography , Heart Rate Variability , Convolution , Transformer
Citation
Yu, S. 2026. Neonatal brain injury detection using advanced deep learning architectures. PhD Thesis, University College Cork.