Learning discriminative time-frequency representations for EEG segment-level seizure-state classification with channel selection

This is the authors' abstract. We don't add key points for this paper.

Abstract Electroencephalography (EEG)-based seizure monitoring supports the diagnosis and management of epilepsy; however, high-dimensional multichannel EEG recordings impose substantial computational burden and limit suitability for portable monitoring systems. This paper proposes a reduced-channel framework for segment-level classification of EEG into ictal, preictal and interictal states. It integrates attention-guided EEG channel selection, high-resolution time-frequency analysis via the Time Reassigned MultiSynchroSqueezing Transform (TMSST) and Concentration of Frequency and Time (ConceFT), and a hybrid CNN-BiLSTM classifier. The framework is evaluated on the CHB-MIT and Siena Scalp EEG databases under two complementary protocols: a stratified 70/30 holdout split and rigorous Leave-One-Subject-Out (LOSO) cross-validation to assess subject independent generalization. Under holdout evaluation using a single EEG channel, the proposed framework achieves an accuracy of 99.28 ± 0.63%, sensitivity of 99.18%, specificity of 99.64%, and MCC of 0.97 on CHB-MIT, and an accuracy of 98.75 ± 0.74%, sensitivity of 98.77%, specificity of 99.38%, and MCC(Matthews correlation coefficients) of 0.97 on Siena. LOSO cross-validation demonstrates subject independent generalization, exceeding 91% accuracy on both databases. These results demonstrate that the framework achieves an effective balance among segment-level classification accuracy, computational efficiency and reduced-channel operation. Moreover, applicability for portable and wearable EEG-based seizure-state monitoring applications is investigated.