Convolutional neural networks ensemble model for neonatal seizure detection

Area of Science:

• Neurology
• Medical Technology
• Artificial Intelligence

Background:

• Neonatal seizures are common and require prompt detection.
• Current methods often rely on manual feature extraction and machine learning.
• Deep learning offers a potential alternative for improved classification.

Purpose of the Study:

• To develop and evaluate a deep learning model for neonatal seizure classification.
• To classify seizure activity directly from raw electroencephalogram (EEG) signals.
• To compare the model's performance against existing methods.

Main Methods:

• A two-dimensional convolutional neural network (2D CNN) architecture was employed.
• The model was trained on expert-annotated EEG data.
• Ten-fold cross-validation was used to assess classification performance.

Main Results:

• Individual models achieved average accuracies (ACC) ranging from 90.1% to 95.6% and average area under the curve (AUC) from 96.7% to 99.2%.
• An ensemble model combining three individual models yielded an average ACC of 96.3% and AUC of 99.3%.
• The study demonstrated superior performance with small time windows (1s).

Conclusions:

• The proposed deep learning approach shows significant promise for accurate neonatal seizure detection.
• This method outperforms previous studies in terms of accuracy and AUC.
• The model is suitable for detecting seizure activity in new clinical data.