Improved attention-based PCNN with GhostNet for epilepsy seizure detection using EEG and fMRI modalities: extractive...

Area of Science:

• Neurology
• Machine Learning
• Medical Imaging

Background:

• Epileptic seizure detection is challenging due to complex EEG signal characteristics.
• Existing machine learning (ML) and deep learning (DL) methods face limitations in interpretability, spatial-temporal modeling, and generalization.

Purpose of the Study:

• To propose an enhanced hybrid parallel convolutional-GhostNet framework (HPG-ESD) for robust seizure detection.
• To leverage multimodal electroencephalogram (EEG) and functional Magnetic Resonance Imaging (fMRI) data for improved detection.

Main Methods:

• Utilized pediatric scalp EEG and resting-state fMRI data from multiple datasets.
• Extracted spatial, temporal, and spectral EEG features with enhanced common spatial pattern (E-CSP).
• Extracted fMRI features using 3D CNN embeddings and smoothened pyramid histogram of oriented gradients (S-PHOG), fused within a soft voting hybrid parallel convolutional-GhostNet (S-HPCGN) model.

Main Results:

• The HPG-ESD framework achieved high performance metrics: 0.941 accuracy, 0.939 precision, and 0.944 sensitivity.
• Outperformed conventional unimodal and state-of-the-art seizure detection methods.

Conclusions:

• Multimodal learning integrating EEG and fMRI shows significant potential for reliable seizure detection.
• Lightweight, attention-enhanced architectures are effective for clinically relevant seizure detection.