Time-frequency embedding with contrastive pre-training allows sub-second seizure detection

Rapid and accurate detection of electrographic seizures is critical for both clinical diagnosis and neuroscience research. Although seizure identification is commonly performed in the time domain, analysis in the time-frequency domain provides a more comprehensive representation of seizure characteristics. In this study, we present a 3D convolutional neural network (CNN) that incorporates a trainable continuous wavelet transform (CWT) layer, enabling adaptive time-frequency feature learning directly from raw EEG. To address common data challenges, we augment the 3D CNN for pre-training with contrastive learning, comparing contrastive predictive coding (CPC) against bidirectional contrastive learning (BiCL). On single-channel and multi-channel data, the standard 3D CNN outperformed both a 2D CNN with pre-computed CWT and a 1D CNN that processes raw signals, achieving > 95% accuracy down to 0.5-second segments. Compared to the standard 3D CNN, the 3D CNN with BiCL pre-training showed superior performance in both low-data and class imbalance scenarios. Further experiments involving band-pass filtering and temporal shuffling revealed that classification is driven primarily by low-frequency patterns and statistical features rather than temporal dependencies. The proposed framework also maintained > 90% accuracy with moderate noise and downsampling applied to inputs, as well as when cross-subject generalization was evaluated using held-out subjects. We show that a 3D CNN with a trainable CWT layer and BiCL pre-training enables accurate sub-second seizure detection and effectively mitigates data limitations common in clinical settings. This work demonstrates that time-frequency embedding within CNNs, augmented by self-supervised pre-training, offers a promising path toward architectures for sub-second seizure detection in the presence of practical limitations of real-world scenarios.