Preictal period optimization for deep learning-based epileptic seizure prediction | JoVE Visualize

Area of Science:

Neuroscience
Biomedical Engineering
Machine Learning

Background:

Accurate seizure prediction is crucial for managing drug-resistant epilepsy and enhancing patient quality of life.
Deep learning models show promise for seizure prediction using electroencephalogram (EEG) signals, but defining the optimal preictal period remains challenging due to state variability.
Current methods struggle to capture the dynamic nature of the preictal state, impacting the reliability of seizure prediction systems.

Purpose of the Study:

To develop novel measures for assessing model behavior under different preictal period definitions.
To propose a data-centric deep learning methodology for identifying the optimal preictal period (OPP) in epilepsy patients.
To evaluate the impact of varying preictal definitions on seizure prediction performance and timing.

Main Methods:

Trained a subject-specific CNN-Transformer model on the CHB-MIT EEG dataset for preictal segment detection.
Fitted sigmoidal curves to model outputs from continuous EEG recordings to analyze prediction dynamics.
Derived performance metrics (convergence, error, stability) and calculated the Continuous Input-Output Performance Ratio (CIOPR) to compare preictal definitions (60, 45, 30, 15 min).

Main Results:

Achieved state-of-the-art performance with the CNN-Transformer model (AUC 99.35%, F1-score 97.46%).
The 60-minute preictal definition yielded earlier predictions, lower preictal error, and reduced output fluctuations, resulting in significantly higher CIOPR scores (p<0.001).
Conventional accuracy metrics were less sensitive to preictal definition variations compared to CIOPR; patient-specific prediction time heterogeneities were observed.

Conclusions:

Novel metrics reveal that preictal period duration significantly impacts prediction timing, a nuance missed by traditional accuracy metrics.
Understanding these impacts and patient-specific variations is vital for developing personalized intelligent epilepsy management systems.
The findings highlight practical limitations in real-world preictal period detection and the need for tailored clinical applications.