Neonatal Seizure Detection Based on Spatiotemporal Feature Decoupling and Domain-Adversarial Learning.

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence

Background:

• Neonatal seizures are key indicators of neurological injury.
• Automated electroencephalogram (EEG) seizure detection faces challenges due to high inter-subject variability.
• A generalization gap exists in current cross-subject seizure detection models.

Purpose of the Study:

• To develop a robust cross-subject seizure detection model for neonatal EEG signals.
• To address the generalization gap caused by inter-subject variability.
• To improve the accuracy and efficiency of automated neonatal seizure detection.

Main Methods:

• A Domain-Adversarial Spatiotemporal Network (DA-STNet) was developed using Short-Time Fourier Transform (STFT) spectrograms.
• The architecture incorporates a Channel-Independent CNN (CI-CNN), Spatial Bidirectional Long Short-Term Memory (Bi-LSTM), and Attention Pooling.
• Domain-adversarial training with a Gradient Reversal Layer (GRL) was employed for domain invariance.

Main Results:

• The DA-STNet achieved state-of-the-art performance with an AUC of 0.9998 and an F1-score of 0.9952 under 5-fold cross-validation.
• Optimal generalization was achieved using only 80% of the source data, demonstrating superior data efficiency.
• The model effectively decoupled pathological features from subject-specific identities.

Conclusions:

• The proposed DA-STNet offers a robust solution for cross-subject neonatal seizure detection.
• The method reduces the need for extensive clinical annotations while maintaining high diagnostic precision.
• This approach shows promise for complex clinical scenarios requiring reliable automated seizure detection.