Epileptic seizure prediction using successive variational mode decomposition and transformers deep learning network

Area of Science:

• Neurology
• Signal Processing
• Machine Learning

Background:

• Epilepsy is a common neurological disorder causing significant patient distress.
• Recurrent seizures necessitate reliable prediction methods for patient care.
• Current prediction methods require enhancement for accuracy and efficiency.

Purpose of the Study:

• To propose a novel epileptic seizure prediction approach.
• To leverage multidimensional successive variational mode decomposition (SVMD) for time-frequency analysis.
• To utilize transformers for enhanced seizure detection.

Main Methods:

• Multivariate SVMD was employed for adaptive decomposition of multi-channel EEG signals into intrinsic modes across different time scales.
• Irrelevant modes were identified and removed through a preprocessing step.
• Denoised data's power spectrum was fed into a pre-trained bidirectional encoder representations from transformers (BERT) for prediction.

Main Results:

• The proposed method demonstrated effective time-frequency analysis of intracranial EEG data.
• The BERT model successfully identified seizure-related mode information.
• Achieved an average sensitivity of 0.86 and a false prediction rate (FPR) of 0.18/h.

Conclusions:

• The combined SVMD and BERT approach offers a promising strategy for epileptic seizure prediction.
• Multidimensional SVMD enhances the analysis of complex EEG signals.
• The method shows potential for improving clinical management of epilepsy through accurate seizure forecasting.