SL2TF achieved state-of-the-art epileptic seizure detection accuracy on benchmark EEG datasets, maintaining strong performance even with limited labeled data.
A novel symmetric time-frequency self-supervised learning framework (SL2TF) improves epileptic seizure detection accuracy in EEG signals, including in low-label scenarios.
Absolute Event Rate: 0% vs 0%
Epileptic seizure detection, a critical task in electroencephalogram (EEG) analysis, faces significant challenges due to the highly non-stationary nature of signal patterns and substantial inter-subject distributional discrepancies. Existing time-domain-centric approaches often fail to comprehensively capture the multi-scale temporal-spectral characteristics of EEG signals during seizures. To address this limitation, we propose SL2TF, a novel self-supervised contrastive learning framework featuring a symmetric time-frequency architecture for robust and efficient seizure detection. The framework first transforms raw EEG signals into the frequency domain via Fast Fourier Transform (FFT), then applies domain-specific data augmentation strategies to generate multi-scale positive-negative sample pairs within both domains. A dual-attention Transformer architecture is employed to learn discriminative representations, while a contrastive auxiliary pre-training task aligns cross-domain features to facilitate collaborative learning. Extensive experiments demonstrate that SL2TF achieves state-of-the-art performance on diverse benchmark datasets. Notably, the framework maintains strong detection accuracy even in low-label scenarios, highlighting its clinical utility. This work underscores the potential of multi-scale time-frequency collaborative self-supervision for complex EEG recognition tasks.
Li et al. (Mon,) reported a other. SL2TF achieved state-of-the-art epileptic seizure detection accuracy on benchmark EEG datasets, maintaining strong performance even with limited labeled data.