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

BACKGROUND AND OBJECTIVE: Epilepsy is a common neurological disorder characterized by recurrent seizures that substantially affect patients' quality of life. Electroencephalography (EEG) is crucial for monitoring brain activity, but traditional spectrogram-based methods often fail to highlight critical frequency bands or capture long-term temporal dynamics. To tackle these challenges, we propose the Epileptic Spectrogram to State-Space (Epi-Spec2State), a convolutional state space model via spectrogram image sequences for prediction-oriented seizure state classification. METHODS: This study employs a private clinical stereo-electroencephalography (SEEG) dataset collected from four patients with drug-refractory epilepsy, and three public EEG datasets including Bonn, CHB-MIT, and Siena. Epi-Spec2State transforms EEG signals into enhanced time-frequency image sequences using short-time Fourier transform and nonlinear frequency mapping to highlight critical bands. It then integrates convolutional layers and pool operation with state space models to capture spatiotemporal dependencies effectively while preserving temporal dynamics through sliding windows. RESULTS: Extensive experiments with respect to patient-specific, mixed-subject, and cross-subject evaluations show the superiority of Epi-Spec2State. Across four patient-specific tasks on the clinical SEEG dataset, the accuracy and specificity reach 96.40%-96.60% and 98.20%-98.30%, respectively. Mixed-subject and cross-subject validation on public datasets further demonstrates its superiority over nine state-of-the-art methods across multiple metrics. CONCLUSION: The proposed Epi-Spec2State can effectively capture the complex spatiotemporal dependencies in EEG signals. Its ability to handle diverse EEG types highlights its potential to support seizure state analysis and pre-seizure warning in clinical practice.