Neonatal seizure detection from raw multi-channel EEG using a fully convolutional architecture.

Area of Science:

• Medical technology
• Artificial intelligence in medicine
• Neonatal neurology

Background:

• Neonatal seizures require accurate detection for timely intervention.
• Current methods often rely on handcrafted features from electroencephalogram (EEG) signals.
• Limited availability of precisely labeled data poses a challenge for traditional machine learning.

Purpose of the Study:

• To develop a deep learning classifier for detecting seizures directly from raw neonatal EEG signals.
• To improve upon the performance and efficiency of existing seizure detection systems.
• To leverage large, weakly labeled datasets for training.

Main Methods:

• Utilized a convolutional neural network (CNN) architecture processing multichannel time-domain EEG signals.
• Trained the model on 834 hours of continuous EEG recordings, exploiting weakly labeled data.
• Validated the system on a public dataset and compared it against Support Vector Machine (SVM) baselines.

Main Results:

• Achieved a 56% relative improvement over a feature-based state-of-the-art baseline.
• Reached an Area Under the Curve (AUC) of 98.5% for seizure detection.
• Demonstrated favorable performance and runtime compared to SVM-based systems.

Conclusions:

• The proposed deep learning architecture enables end-to-end optimization for neonatal seizure detection.
• This approach enhances the efficient use of available training data, reducing reliance on precise clinical labels.
• Opens new possibilities for applying deep learning in neonatal EEG analysis.