Patient-Specific Seizure Prediction via Adder Network and Supervised Contrastive Learning.

Area of Science:

• Computational Neuroscience
• Machine Learning for Healthcare

Background:

• Deep learning (DL) methods are prevalent for seizure prediction from electroencephalogram (EEG).
• Existing DL models often suffer from high computational complexity due to numerous multiplication operations.
• Current approaches frequently overlook intrinsic data patterns, focusing instead on specialized architectures.

Purpose of the Study:

• To propose a computationally efficient and effective end-to-end deep learning model for seizure prediction.
• To reduce computational cost by replacing multiplication with addition in convolutional processes.
• To leverage supervised contrastive learning for enhanced representation learning and accurate seizure detection.

Main Methods:

• Introduced an Adder Network with Supervised Contrastive Learning (AddNet-SCL), utilizing addition instead of multiplication.
• Employed contrastive learning to cluster similar data points and separate dissimilar ones in the projection space.
• Combined supervised contrastive loss and cross-entropy loss for model training, with an adaptive learning rate strategy.

Main Results:

• Achieved 94.9% sensitivity, 94.2% AUC, and 0.077/h FPR on the CHB-MIT database.
• Obtained 89.1% sensitivity, 83.1% AUC, and 0.120/h FPR on the Kaggle database.
• Demonstrated competitive performance compared to existing seizure prediction methods.

Conclusions:

• The proposed AddNet-SCL method offers a significant reduction in computational complexity for seizure prediction.
• The model effectively utilizes intrinsic data patterns and label information through contrastive learning.
• The achieved results indicate broad prospects for clinical application in seizure prediction.