Epileptic EEG Classification by Using Time-Frequency Images for Deep Learning.

Area of Science:

• Neurology
• Biomedical Engineering
• Signal Processing

Background:

• Epilepsy is a common neurological disorder affecting millions globally.
• Electroencephalogram (EEG) recordings are crucial for epilepsy diagnosis and patient monitoring.
• Existing computer-aided diagnosis systems for seizure detection and prediction have limitations.

Purpose of the Study:

• To propose a novel method for detecting and predicting epileptic seizures using EEG signals.
• To leverage the high-resolution time-frequency representation capabilities of Synchrosqueezing Transform (SST).
• To integrate SST with Convolutional Neural Networks (CNN) for enhanced seizure analysis.

Main Methods:

• Utilized Fourier-based Synchrosqueezing Transform (SST) for high-resolution time-frequency analysis of EEG signals.
• Developed a Convolutional Neural Network (CNN) model incorporating SST features for seizure detection and prediction.
• Evaluated the proposed SST-based CNN method on two datasets: the collected IKCU dataset and the public CHB-MIT dataset.

Main Results:

• The SST-based CNN method achieved high segment-based seizure detection precision and accuracy on both datasets.
• IKCU dataset results: 98.99% precision and 99.06% accuracy.
• CHB-MIT dataset results: 99.81% precision and 99.63% accuracy.
• Significantly improved segment-based seizure prediction performance on the CHB-MIT dataset (98.54% precision, 97.92% accuracy) compared to existing methods.

Conclusions:

• The proposed SST-based CNN approach effectively detects and predicts epileptic seizures from EEG signals.
• SST's ability to localize energy in the time-frequency plane is beneficial for analyzing sudden epileptic discharges.
• This novel method demonstrates superior performance and holds promise for clinical application in epilepsy management.