A Contrastive Learning-Enhanced Residual Network for Predicting Epileptic Seizures Using EEG Signals

Area of Science:

• Computational Neuroscience
• Machine Learning for Healthcare
• Biomedical Signal Processing

Background:

• Epileptic seizure prediction models face challenges with large labeled datasets and complex EEG data.
• Current models struggle with robustness and generalization due to data limitations.

Purpose of the Study:

• To propose CLResNet, a novel framework combining contrastive self-supervised learning and a modified deep residual neural network.
• To address the limitations of traditional models by reducing reliance on labeled data and enhancing predictive accuracy.

Main Methods:

• Utilized contrastive self-supervised learning (CL) for pre-training on unlabeled EEG data to extract robust features.
• Employed a modified deep residual neural network (ResNet) architecture with residual connections for efficient gradient flow and deep feature learning.
• Fine-tuned the model on smaller labeled datasets to improve efficiency and predictive performance.

Main Results:

• CLResNet achieved high accuracy (92.97%) and sensitivity (94.18%) on the CHB-MIT dataset, outperforming existing methods.
• Demonstrated competitive performance on the Siena dataset with 92.79% accuracy and 91.47% sensitivity.
• Showcased a low false-positive rate (0.043/h on CHB-MIT, 0.041/h on Siena), indicating high reliability.

Conclusions:

• CLResNet effectively addresses EEG data variations, proving contrastive self-supervised learning as a robust approach for seizure prediction.
• The framework significantly enhances model robustness and generalizability by leveraging unlabeled data.
• The study highlights the potential of CLResNet for accurate and efficient epileptic seizure prediction.