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

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) analysis is crucial for diagnosing seizures in clinical settings and advancing neuroscience research.
• Traditional time-domain analysis of EEG data for seizure detection offers limited insights.
• Time-frequency domain analysis provides a more comprehensive understanding of seizure characteristics.

Purpose of the Study:

• To develop a novel 3D convolutional neural network (CNN) capable of adaptive time-frequency feature learning from raw EEG data.
• To investigate the efficacy of contrastive learning strategies, specifically contrastive predictive coding (CPC) and bidirectional contrastive learning (BiCL), for pre-training the 3D CNN.
• To evaluate the performance of the proposed framework in detecting electrographic seizures with high accuracy and robustness.

Main Methods:

• A 3D CNN architecture was designed, integrating a trainable continuous wavelet transform (CWT) layer for direct time-frequency feature extraction from raw EEG signals.
• Contrastive learning techniques (CPC and BiCL) were employed for pre-training the 3D CNN to address challenges like limited data and class imbalance.
• The model's performance was assessed using single-channel and multi-channel EEG data, including evaluations under noisy conditions, downsampling, and cross-subject generalization.

Main Results:

• The 3D CNN with a trainable CWT layer significantly outperformed 2D CNN and 1D CNN models, achieving over 95% accuracy for seizure detection in segments as short as 0.5 seconds.
• Contrastive pre-training enhanced the 3D CNN's performance, particularly in scenarios with limited data and imbalanced classes.
• The model demonstrated robustness, maintaining over 90% accuracy with moderate noise, downsampling, and when generalizing to unseen subjects, indicating that low-frequency patterns and statistical features are key drivers of classification.

Conclusions:

• A 3D CNN incorporating a trainable CWT layer and contrastive pre-training enables accurate sub-second seizure detection, effectively addressing data limitations prevalent in clinical EEG analysis.
• This approach offers a promising direction for developing robust seizure detection architectures that can handle real-world data challenges.
• Time-frequency embedding within CNNs, bolstered by self-supervised pre-training, represents a significant advancement for sub-second seizure detection systems.