Deep learning approaches for diagnosing seizure based on EEG signal analysis

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy affects 1% of the global population, necessitating improved seizure detection methods.
• Current deep learning (DL) models for seizure detection face challenges like patient variability and noisy EEG signals.
• Existing methods struggle with generalization and require enhanced robustness.

Purpose of the Study:

• To develop a personalized seizure detection system using an Ensemble of Deep Transfer Learning (EDTL) models.
• To improve the accuracy and generalization of seizure detection by combining multiple DL architectures.
• To address limitations of current methods, including inter-patient variability and signal noise.

Main Methods:

• EEG data from epilepsy patients was transformed into spectrograms using personalized sliding windows and Short Time Fourier Transform (STFT).
• A novel EDTL framework integrated pre-trained models (ResNet, EfficientNet) with a custom 2D Convolutional Neural Network (2DCNN).
• Models were trained independently on patient-specific features and then combined to enhance robustness and adaptability.

Main Results:

• The EDTL model demonstrated superior performance compared to individual models across standard metrics.
• Achieved a high Area Under the Curve (AUC) of 99.23% on the CHB-MIT and Turkish Epilepsy EEG datasets.
• The EDTL framework showed improved robustness against noise and better adaptability to patient-specific EEG variations.

Conclusions:

• The proposed EDTL model offers a significant advancement in personalized epilepsy seizure detection.
• This approach effectively overcomes limitations of single DL models, providing more reliable seizure prediction.
• EDTL enhances generalization and adaptability, paving the way for improved clinical applications in epilepsy management.