Machine Learning Algorithm for Epileptic Seizure Prediction from Scalp EEG Records

Epileptic seizures need to be predicted with sufficient time to allow the patients to prevent clinical symptoms by taking their prescribed anti-epileptic medications or via neurostimulation. In this context, automatic seizure prediction techniques with high sensitivity and specificity are needed. In this work, we present a machine learning method to predict epileptic seizures by analyzing the pre-ictal stage of scalp EEG recordings. The TUH EEG Seizure Corpus files were used to train and test the technique. After filtering and channel normalization, statistical parameters from the Teager energy operator, the absolute band power, the Hjorth parameters, and the kurtosis of the signal were extracted from each channel. Four machine learning classifiers were implemented after a feature selection step based on inter-feature cross-correlation values: Random Forest, XGBoost, Support Vector Machine, and K-nearest neighbors. The XGBoost model yielded the best performance metrics with the validation set: 99.84% in accuracy, 100% in precision, 99.6% in recall, 100% in specificity, and 0.998 in the area under the receiving operating characteristics curve (AUC). Ultimately, these methods can detect subtle changes in scalp EEG records that forecast an epileptic seizure.