Explainable Automated Seizure Detection using Attentive Deep Multi-View Networks.

Area of Science:

• Neuroscience and Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Manual Electroencephalography (EEG) analysis for epileptic seizure detection is labor-intensive and suffers from variability.
• EEG signals are often corrupted by noise and artifacts, complicating accurate seizure identification.
• Existing multi-view seizure detection systems lack dynamic weighting for optimal feature contribution.

Purpose of the Study:

• To develop an interpretable deep learning model for enhanced epileptic seizure detection.
• To address the challenge of dynamic weight assignment in multi-view seizure detection systems.
• To improve the accuracy and reliability of seizure detection from noisy EEG data.

Main Methods:

• Proposed a fusion attentive deep multi-view network (fAttNet) incorporating temporal multi-channel EEG, wavelet packet decomposition (WPD), and hand-engineered features.
• Implemented an artifact rejection approach to filter non-brain signal noise.
• Employed an attention mechanism for dynamic weighting of different data views.

Main Results:

• The fAttNet model achieved improved performance on the Temple University Hospital (TUH) seizure database.
• Accuracy increased from 0.82 to 0.86, and F1-score improved from 0.78 to 0.81 compared to state-of-the-art methods.
• The model demonstrated interpretability, aiding clinicians in identifying seizure-affected brain regions.

Conclusions:

• The proposed fAttNet offers a more accurate and interpretable solution for epileptic seizure detection using EEG.
• Dynamic weighting and artifact rejection significantly enhance detection performance.
• The interpretability of fAttNet supports clinical decision-making in epilepsy management.