Hypoxic-ischemic encephalopathy grading using novel EEG signal processing and machine learning techniques

\IEEEPARstart{H}{ypoxic-Ischemic Encephalopathy} (HIE) is a critical brain injury in newborns resulting from oxygen or blood supply deprivation during birth. Traditional diagnosis methods for HIE require specialized expertise from a trained medical professional, and immediate intervention of treatment is required within the first 6 hours post primary hypoxic-ischemic insult. Electroencephalogram (EEG) is the gold standard for analysing brain pathologies for neonates and large windows of this complex signal must be assessed for accurate HIE diagnosis. Using an openly available dataset, this thesis proposes an automated approach for HIE grading, leveraging signal processing techniques and Artificial Intelligence (AI) to provide accurate and timely assessments, with an implementation enabling edge-device deployment for real-world clinical utility. Representing the 1 hour epoch of the EEG signal in the amplitude and frequency domain through Mel Spectorgram representation, the HIE grading task becomes an image recognition problem, where Convolutional Neural Networks have shown high accuracy and efficiency. An initial test accuracy of 84.85\% is achieved. Further enhancement of the signals rhythmic behaviour is necessary to increase the grading potential of the signal, thus the FM/AM sonification algorithm was implemented, transforming the EEG signal into an amplitude and frequency modulated audio signal. This pre-processing is designed to enhance the background rhythmic pattern of the signal, an essential feature used by the clinician for visual inspection. The two dimensional (2D) CNN is designed as a regressor to map the input image to a value on the HIE grading scale to enhance the model's ability to leverage the monotonic relationship between the grades. An optimised rounding function is implemented to define the final clinical grade as a novel post-processing technique. An overall test accuracy of 89.97\% based on a rigourous nested cross-validation evaluation framework is achieved, surpassing the current state-of-the-art by 6\%. The robustness and generalisability of the proposed method is obvious from the weighted F1-score of 0.8985 and a Kappa score of 0.8219. While enhancing the accuracy and speed of HIE diagnosis with this AI-driven approach, the goal is to make it readily available at the point of care. An inference time of 62.7 milliseconds is achieved for the quantized model on the Snapdragon processor highlighting its suitability for real-time, on-device HIE grading.

Grantor dc:publisher

University College Cork

Year dc:date.issued

2025

Author dc:creator

• Twomey, Leah

Advisors dc:contributor.advisor

• Popovici, Emanuel
• Temko, Andriy

• Machine learning
• EEG
• AI
• Sonification
• Hypoxic-ischemic enchephalopathy
• Neural networks

Statement dc:rights

• © 2025 , Leah Twomey.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-nc/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/17533

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/17533