[Automatic epilepsy detection with an attention-based multiscale residual network].

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Context:

• Epilepsy diagnosis relies heavily on electroencephalogram (EEG) interpretation, which is time-consuming and prone to human error.
• Deep learning models offer a promising avenue for automating EEG analysis, enhancing diagnostic efficiency and accuracy.
• Understanding the complex spatio-temporal dynamics and inter-channel information flow in epileptic EEG is crucial for effective detection.

Purpose:

• To propose and evaluate an attention-based multi-scale residual network (AMSRN) for the automatic detection of epilepsy from EEG signals.
• To leverage multiscale principal component analysis (MSPCA) for noise reduction and feature enhancement of raw EEG data.
• To design and implement a novel deep learning architecture that effectively captures multi-scale and spatio-temporal features in epilepsy EEG.

Summary:

• The proposed AMSRN model integrates multiscale principal component analysis (MSPCA) for preprocessing, followed by an attention module (AM), multiscale convolutional module (MCM), spatio-temporal feature extraction module (STFEM), and classification module (CM).
• MSPCA enhances feature extraction by reducing noise and improving signal quality.
• The AMSRN architecture utilizes an attention-weighted mechanism to re-express signals and effectively fuses multi-scale and spatio-temporal features for classification.

Impact:

• The AMSRN model achieved high performance on the CHB-MIT dataset, demonstrating excellent sensitivity (98.56%), F1 score (98.35%), accuracy (98.41%), and precision (98.43%).
• The model effectively utilizes brain network information flow during seizures, enhancing inter-channel differentiation.
• This deep learning approach significantly improves the performance of automatic epilepsy detection by capturing complex EEG characteristics.