Towards Automated EEG-Based Epilepsy Detection Using Deep Convolutional Autoencoders.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Epilepsy is a common neurological disorder necessitating accurate seizure detection.
• Manual analysis of electroencephalography (EEG) for seizures is time-intensive and requires expertise.
• Current automated methods, including deep learning, face challenges in sensitivity, false alarm rates, and optimal EEG data representation.

Purpose of the Study:

• To develop an improved deep learning approach for seizure detection using EEG.
• To investigate the effectiveness of a Deep Convolutional Autoencoder (DCAE) for extracting essential EEG features.
• To evaluate whether combining time and frequency domain information enhances EEG feature representation for epilepsy monitoring.

Main Methods:

• Proposed a Deep Convolutional Autoencoder (DCAE) model for extracting low-dimensional latent representations from EEG signals.
• Trained and evaluated multiple autoencoders using different loss functions based on time and frequency domains.
• Assessed the model's ability to preserve relevant information by comparing reconstruction errors between time series and frequency-domain representations.

Main Results:

• The DCAE model incorporating both time series and frequency domain losses demonstrated superior reconstruction performance.
• This suggests that relying on a single data representation (time or frequency) may not adequately preserve critical EEG signal properties.
• The study provides insights into deep learning processing of EEG and the capture of frequency information from time series data.

Conclusions:

• A DCAE integrating both time and frequency domain losses offers a more effective method for EEG feature extraction in epilepsy research.
• Deep learning models may benefit from multi-domain input representations for comprehensive EEG signal analysis.
• This approach advances automated seizure detection by improving the preservation of relevant EEG signal characteristics.