Automatic detection of epileptic seizure based on one dimensional cascaded convolutional autoencoder with adaptive...

Area of Science:

• Neurology
• Machine Learning
• Biomedical Signal Processing

Background:

• Identifying seizure occurrence periods (SOP) in extended EEG recordings is critical for epilepsy diagnosis.
• Existing computer-aided diagnosis systems for epileptic seizure detection (ESD) often require labeled data and focus on ictal/interictal states, limiting clinical application.
• Supervised learning approaches for ESD necessitate extensive manual data labeling, posing a significant bottleneck.

Purpose of the Study:

• To develop an unsupervised learning framework for epileptic seizure detection (ESD) in long-term EEG recordings.
• To enable accurate identification of seizure occurrence periods (SOP) without the need for labeled data.
• To create a robust and adaptable automated system for clinical use in EEG analysis.

Main Methods:

• Utilized a 1D-cascaded convolutional autoencoder (1D-CasCAE) for unsupervised learning on EEG data.
• Segmented EEG recordings into 5-second epochs and selected eight informative channels based on correlation and entropy.
• Employed adaptive thresholding and a moving window approach to enhance the model's robustness in detecting anomalies (ictal segments).

Main Results:

• The 1D-CasCAE model effectively learned normal EEG patterns and identified anomalies (ictal segments) using reconstruction errors.
• Achieved superior performance compared to other anomaly detection methods, with average Gmean of 98.00% and sensitivity of 94.94% on CHB-MIT datasets.
• Demonstrated high specificity (99.60%) and a low false positive rate (0.0044 h⁻¹) for typical patients.

Conclusions:

• The proposed unsupervised 1D-CasCAE framework provides an effective method for automated epileptic seizure detection (ESD).
• This model can replace manual EEG inspection by neurologists, improving efficiency in clinical settings.
• The system's adaptability with variable time windows allows for patient-specific seizure occurrence period (SOP) detection.