Transformer-Based Wavelet-Scalogram Deep Learning for Improved Seizure Pattern Recognition in Post-Hypoxic-Ischemic...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Neonatal Medicine

Background:

• Hypoxic-ischemic (HI) events in newborns can lead to seizures and neurodevelopmental impairments.
• Accurate seizure detection is challenging, necessitating automated diagnostic methods.
• Previous deep learning models showed promise in identifying post-HI seizures in fetal sheep.

Purpose of the Study:

• To evaluate the effectiveness of Transformer models for detecting seizures in fetal sheep EEG after HI.
• To compare Transformer model performance against prior deep convolutional neural network (CNN) methods.
• To assess the potential of these models as clinical decision support tools for neonatal seizures.

Main Methods:

• Transformer models were trained on wavelet scalogram (WS) images of EEG patterns from fetal sheep recovering from HI.
• A subset of 800 WS images of seizure and non-seizure EEG patterns was used for training.
• The Visual Transformer (ViT) architecture was specifically assessed.

Main Results:

• The Visual Transformer (ViT) achieved a high overall accuracy of 99.5% with an AUC of 0.995.
• ViT performance surpassed that of previous deep CNN models.
• Transformer models demonstrated superior efficiency and robustness in seizure detection.

Conclusions:

• Transformer models, particularly ViT, are highly effective for detecting HI-induced EEG seizures in a preclinical model.
• These models show significant potential as automated clinical decision support tools for neonatal seizure identification.
• The study highlights the advancement of AI in neonatal neurology for improved patient outcomes.