Application of deconvolutional networks for feature interpretability in epilepsy detection

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Scalp electroencephalography (EEG) is crucial for epilepsy diagnosis.
• Current automated seizure detection algorithms often overlook individual channel contributions.
• Fully Convolutional Networks (FCNs) offer interpretability but are underutilized in epilepsy detection.

Purpose of the Study:

• To develop a novel deep learning model for improved patient-independent epilepsy detection.
• To enhance the interpretability of automated seizure detection algorithms.
• To evaluate the contribution of different EEG channels in seizure detection.

Main Methods:

• A novel Convolutional Neural Network (CNN) model was developed, integrating Squeeze-and-Excitation (SE) modules with an FCN architecture.
• The model's patient-independent epilepsy detection performance was assessed using the CHB-MIT dataset.
• Comparative analysis was performed by replacing SE modules with Inception, ResNet, and CBAM modules.

Main Results:

• The proposed SE-FCN model demonstrated superior performance, stability, and reliability compared to alternative configurations.
• Achieved a G-Mean of 82.7% for sensitivity and specificity on the CHB-MIT dataset.
• Quantified channel contributions, identifying FZ, CZ, PZ, FT9, FT10, and T8 as critical brain regions for seizure detection.

Conclusions:

• The novel SE-FCN algorithm offers an accurate and interpretable solution for automated epilepsy detection.
• The study highlights the importance of considering individual channel contributions for robust seizure identification.
• Findings provide insights into key brain regions involved in epileptic seizures, aiding clinical interpretation.