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

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Accurate electroencephalogram (EEG) seizure detection is crucial for clinical diagnosis and research.
• Time-frequency domain analysis offers richer insights into seizure dynamics than traditional time-domain methods.
• Existing methods face challenges with data limitations like noise, downsampling, and class imbalance.

Purpose of the Study:

• To develop and evaluate a novel 3D convolutional neural network (CNN) with an integrated trainable continuous wavelet transform (CWT) layer for adaptive time-frequency feature learning from raw EEG.
• To investigate the efficacy of self-supervised pre-training strategies, specifically contrastive predictive coding (CPC) and bidirectional contrastive learning (BiCL), to enhance CNN performance.
• To assess the framework's robustness against common data challenges, including low data availability, class imbalance, noise, downsampling, and cross-subject generalization.

Main Methods:

• A 3D CNN architecture was designed, incorporating a trainable CWT layer for direct time-frequency feature extraction from EEG signals.
• Contrastive learning techniques, CPC and BiCL, were employed for pre-training the 3D CNN to improve feature representation.
• Performance was evaluated on single-channel and multi-channel EEG data, comparing against 2D CNN and 1D CNN models, and tested under various data-degrading conditions.

Main Results:

• The proposed 3D CNN with a trainable CWT layer achieved over 95% accuracy for seizure detection in segments as short as 0.5 seconds.
• The 3D CNN with BiCL pre-training demonstrated superior performance, particularly in low-data and class-imbalanced scenarios, outperforming the standard 3D CNN.
• The model maintained high accuracy (>90%) even with moderate noise, downsampling, and when generalizing to unseen subjects, indicating robustness.

Conclusions:

• A 3D CNN framework with a trainable CWT layer and BiCL pre-training enables highly accurate, sub-second electroencephalogram seizure detection.
• This approach effectively addresses practical data limitations encountered in clinical settings, offering a robust solution.
• Integrating time-frequency embedding within CNNs, augmented by self-supervised pre-training, presents a promising direction for advanced seizure detection systems.