Automatic Seizure Detection using Hierarchical Spectral-Temporal Feature Learning with an Imbalance-Aware Transformer

Area of Science:

• Neurology
• Computational Neuroscience
• Machine Learning

Background:

• Epilepsy is a chronic neurological disorder characterized by recurrent seizures.
• Electroencephalography (EEG) signal analysis offers potential for automated seizure detection.
• Automated seizure detection faces challenges in feature representation and class imbalance.

Purpose of the Study:

• To develop a novel deep learning architecture for automated seizure detection in EEG signals.
• To address comprehensive feature representation and class distribution imbalance in EEG data.
• To create a practical diagnostic tool for epilepsy management.

Main Methods:

• A multibranch neural network processes EEG signals at various spectral and temporal resolutions.
• An attention-based mechanism refines features, emphasizing clinically relevant characteristics.
• A modified loss function with class-specific margin adjustments handles data imbalance.

Main Results:

• Scalp EEG analysis achieved 96.06% sensitivity and 98.50% specificity with a low FDR of 0.34/h.
• Intracranial EEG analysis showed similar efficacy (95.90% sensitivity, 98.65% specificity) with reduced FDR (0.18/h).
• The model demonstrated consistent performance across diverse EEG modalities.

Conclusions:

• The novel deep learning architecture effectively detects seizures in EEG signals.
• The approach successfully addresses feature representation and class imbalance challenges.
• Validated on scalp and intracranial EEG, the tool shows significant clinical utility for epilepsy diagnosis.