Tiny Convolutional Neural Network with Supervised Contrastive Learning for Epileptic Seizure Prediction.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automatic seizure prediction using ElectroEncephaloGraphy (EEG) is crucial for epilepsy patient safety and anxiety reduction.
• Current methods face performance bottlenecks, patient-specific efficacy variations, and complex model structures.

Purpose of the Study:

• To introduce novel Siamese Network (SiaNet) and Triplet Network (TriNet) models for improved seizure prediction.
• To address limitations of existing methods by developing computationally efficient and accurate models.

Main Methods:

• Applied Short-Time Fourier Transform (STFT) to pre-processed EEG data.
• Utilized tiny convolutional neural networks with supervised contrastive learning to train SiaNet and TriNet.
• Constructed data tuples for training, minimizing intra-class sample intervals and maximizing inter-class intervals.

Main Results:

• Achieved promising seizure prediction results on the CHB-MIT and Siena datasets, involving 35 patients.
• Demonstrated that both SiaNet and TriNet models possess a minimal parameter count of only 19.351K.
• The proposed networks learn effectively through shared weights and contrastive learning across multiple branches.

Conclusions:

• SiaNet and TriNet offer a significant advancement in automatic seizure prediction technology.
• The models provide a more efficient and potentially more generalizable approach compared to existing methods.
• These findings suggest a promising direction for developing practical and effective epilepsy management tools.