A Class-Imbalance Aware and Explainable Spatio-Temporal Graph Attention Network for Neonatal Seizure Detection

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Technology

Background:

• Neonatal seizures are the most common neurological disorder symptom in newborns.
• Accurate and timely seizure detection is critical for effective treatment and improved outcomes.
• Existing methods face challenges with data complexity and class imbalance.

Purpose of the Study:

• To develop an accurate and explainable deep learning model for automated neonatal seizure detection.
• To integrate temporal EEG features with spatial channel information for enhanced detection.
• To address the challenge of class imbalance in neonatal seizure datasets.

Main Methods:

• A novel Spatio-Temporal Graph Attention Network (ST-GAT) combining Convolutional Neural Networks (CNNs) and Graph Attention Networks (GATs).
• Utilized 1D CNNs for temporal feature extraction from EEG segments.
• Employed GATs to model spatial relationships and attention between EEG channels, visualizing important brain regions.
• Incorporated under-sampling and focal loss to handle severe class imbalance.

Main Results:

• The ST-GAT model achieved a mean Area Under the Curve (AUC) of 96.6% and a Kappa score of 0.88.
• Demonstrated superior performance compared to previous benchmarked methods.
• GAT coefficients provided insights into seizure localization by highlighting critical channel interactions.

Conclusions:

• The proposed ST-GAT model offers a highly accurate and explainable approach for automated neonatal seizure detection.
• The method effectively integrates spatio-temporal EEG data and addresses class imbalance.
• ST-GAT shows significant potential for clinical application in neonatal intensive care units.