The interpretable deep learning framework and validation for seizure detection in pediatric electroencephalography:...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptic seizure detection is crucial for clinical intervention.
• Current deep learning models for electroencephalography (EEG) analysis often function as black boxes, limiting clinical interpretability.
• Existing Convolutional Neural Networks (CNNs) and Sequence Generation Networks (SGNs) have limitations in hierarchical modeling, quantifying channel contributions, and capturing complex spatiotemporal dependencies.

Purpose of the Study:

• To propose an interpretable deep learning framework for epileptic seizure detection.
• To compare the performance of two novel models: a Fully Convolutional Network with Squeeze-and-Excitation modules (SE-FCN) and a transformer-based model (TransNet).
• To generate channel saliency weights and heatmaps for inferring potential epileptogenic zones.

Main Methods:

• Developed a SE-FCN model to enhance spatial sensitivity and retain temporal resolution in EEG data.
• Developed a TransNet model utilizing self-attention to capture temporal and channel-wise dependencies.
• Evaluated models on the CHB-MIT pediatric EEG dataset using a subject-independent cross-validation protocol.

Main Results:

• SE-FCN achieved an Area Under the Curve (AUC) of 0.89 and 86.7% accuracy.
• TransNet achieved an AUC of 0.92 and 86.4% accuracy.
• Saliency maps from both models showed high consistency, enabling the categorization of 22 patients into five groups based on inferred seizure origins.

Conclusions:

• Both SE-FCN and TransNet offer interpretable insights into epileptic seizure localization from EEG data.
• The proposed models demonstrate effectiveness in identifying potential epileptogenic zones, outperforming traditional black-box approaches.
• The saliency mapping capability facilitates a better understanding of seizure origins and patient stratification.