Epi-Spec2State: Convolutional state space model via spectrogram image sequences for prediction-oriented seizure state...

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Epilepsy is a neurological disorder causing recurrent seizures, impacting quality of life.
• Traditional electroencephalography (EEG) analysis struggles with critical frequency bands and long-term dynamics.
• Accurate seizure state classification is vital for patient management.

Purpose of the Study:

• To introduce Epileptic Spectrogram to State-Space (Epi-Spec2State), a novel model for seizure state classification.
• To overcome limitations of traditional spectrogram-based EEG analysis.
• To enhance prediction-oriented seizure detection using convolutional state space models.

Main Methods:

• Utilized private SEEG and public EEG datasets (Bonn, CHB-MIT, Siena).
• Transformed EEG signals into enhanced time-frequency images using STFT and nonlinear frequency mapping.
• Integrated convolutional layers, pooling, and state space models with sliding windows for spatiotemporal analysis.

Main Results:

• Epi-Spec2State demonstrated superior performance across patient-specific, mixed-subject, and cross-subject evaluations.
• Achieved high accuracy (96.40%-96.60%) and specificity (98.20%-98.30%) on a clinical SEEG dataset.
• Outperformed nine state-of-the-art methods on public EEG datasets across multiple metrics.

Conclusions:

• Epi-Spec2State effectively captures complex spatiotemporal dependencies in EEG signals.
• The model's versatility across diverse EEG types shows promise for clinical applications.
• Potential for supporting seizure state analysis and pre-seizure warning systems.