FusionXNet: enhancing EEG-based seizure prediction with integrated convolutional and Transformer architectures.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Effective seizure prediction is crucial for reducing patient burden and healthcare costs.
• Current deep learning methods for seizure prediction often use single models, limiting feature extraction capabilities.

Purpose of the Study:

• To develop a hybrid deep learning model, FusionXNet, integrating Convolutional Neural Networks (CNNs) and Transformer architectures for enhanced seizure prediction.
• To improve feature representation by combining local and global electroencephalography (EEG) features.

Main Methods:

• FusionXNet utilizes a token synthesis unit for local feature extraction via convolution.
• Global EEG representations are captured using attention mechanisms.
• Local and global features are merged to enhance representations before classification.

Main Results:

• FusionXNet was evaluated on the Boston Children's Hospital and MIT dataset.
• In event-based, patient-specific experiments, the model achieved 97.602% sensitivity and a 0.059 h^-1 false positive rate (FPR).
• The proposed model outperformed existing methods in seizure prediction accuracy.

Conclusions:

• FusionXNet offers a robust and efficient approach to seizure prediction by effectively combining local and global feature extraction.
• The model's high sensitivity and low FPR suggest significant potential for real-world clinical applications.
• This advancement could lead to improved patient management and reduced healthcare expenditures.