A ResNet-LSTM hybrid model for predicting epileptic seizures using a pretrained model with supervised contrastive...

Area of Science:

• Computational Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Epileptic seizures pose significant challenges due to the complexity of electroencephalography (EEG) data.
• Accurate seizure prediction is crucial for improving patient quality of life and treatment efficacy.

Purpose of the Study:

• To develop and validate a novel deep learning framework for predicting epileptic seizures.
• To leverage supervised contrastive learning and a hybrid ResNet-LSTM model for enhanced prediction accuracy.

Main Methods:

• EEG data transformed into spectrogram images using Short-Time Fourier Transform (STFT).
• Pre-training a Residual Network (ResNet) using supervised contrastive learning on augmented data.
• Developing a hybrid model combining pre-trained ResNet and Long Short-Term Memory (LSTM) for seizure prediction.

Main Results:

• The proposed method achieved high accuracy and sensitivity on CHB-MIT (91.90% accuracy, 89.64% sensitivity) and SNUH (83.37% accuracy, 79.89% sensitivity) datasets.
• Demonstrated superior performance compared to conventional seizure prediction methods.
• Leave-one-out cross-validation confirmed the model's generalization ability.

Conclusions:

• The hybrid ResNet-LSTM model with supervised contrastive learning offers a promising approach for accurate epileptic seizure prediction.
• The STFT-based spectrogram representation effectively captures crucial time-frequency information from EEG signals.
• This method holds potential for real-world clinical applications in epilepsy management.