Probabilistic prediction of Epileptic Seizures using SVM

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy seizure prediction remains a significant clinical challenge.
• Accurate prediction of seizures from intracranial electroencephalography (iEEG) signals is crucial for improving patient care.
• Existing methods often struggle with imbalanced datasets, where non-seizure events heavily outweigh seizure occurrences.

Purpose of the Study:

• To develop and evaluate an efficient algorithm for classifying iEEG signals as ictal (seizure) or interictal (non-seizure).
• To enable accurate and timely prediction of rare seizure events using machine learning.
• To address the challenge of imbalanced class distribution in EEG datasets for seizure prediction.

Main Methods:

• Implementation of a linear Support Vector Machine (SVM) classifier.
• Extraction of various univariate linear measures from iEEG signals.
• Training and testing the model on a dataset of 140 hours of iEEG recordings from 6 patients, including 34 seizures.
• Utilizing a 2-second window and 10-fold cross-validation for performance evaluation.

Main Results:

• The developed SVM classifier demonstrated strong performance despite a highly imbalanced dataset.
• The algorithm achieved a sensitivity of 78% and a specificity of 100%.
• The model successfully predicted some seizures with up to 0.4 probability, 30-40 minutes in advance.

Conclusions:

• The proposed linear SVM algorithm offers an efficient and accurate method for human seizure prediction from iEEG data.
• The approach effectively handles imbalanced datasets, crucial for reliable seizure forecasting.
• This work contributes to advancing the capabilities of automated seizure detection and prediction systems.