Epileptic seizure detection by using interpretable machine learning models.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Epileptic seizure detection from electroencephalogram (EEG) data is crucial for diagnosis but visually intensive for experts.
• Current seizure detection methods often lack interpretability and struggle with imbalanced datasets.
• There is a need for automated, efficient, and explainable seizure detection systems.

Purpose of the Study:

• To develop an automated seizure detection system for EEG data that mimics clinical expert visual diagnosis.
• To enhance seizure detection by providing explainable AI (XAI) insights into the model's decision-making process.
• To address dataset imbalance issues common in clinical EEG data.

Main Methods:

• EEG data was processed as images, utilizing established deep learning models like LeNet, VGG, ResNet, and Vision Transformer (ViT).
• A novel data augmentation technique, Random Channel Ordering (RCO), was introduced to balance seizure and non-seizure classes.
• Model interpretability was achieved using Gradient-weighted Class Activation Mapping (Grad-CAM) and attention layer methods.

Main Results:

• The RCO data augmentation method effectively balanced the dataset, improving model performance.
• Deep learning models demonstrated strong performance in classifying seizure and non-seizure EEG images.
• Grad-CAM and attention mechanisms provided clear explanations for model detections and quantified seizure severity.

Conclusions:

• Processing EEG as images allows flexible application of diverse machine learning models.
• The RCO method offers a robust solution to clinical data imbalance.
• The developed system provides intuitive, interpretable results for clinical experts, serving as a valuable diagnostic reference.