Subject-Independent Epileptic Seizure Prediction using Spectral Power and Correlation Coefficients

Sammanfattning: Epileptic seizure prediction algorithms with prediction rates above random have been produced, with varying success, during the last ten to twenty years. The algorithms produced have been tailored to the specific characteristics of a subject’s epilepsy, referred to as subject-specific prediction algorithms. Such customization entails the training of the algorithm’s classifier on the specific EEG-data pertaining to the subject. An inherent requirement is the time-intensive task of recording and labeling the subjects EEG, which will be used for the training of the classifier. As such, this thesis investigates the possibility of adjusting the training of a subject-specific algorithm’s classifier to make it subject-independent. The investigation is based on whether the subject-independent version could achieve prediction rates equal to or better than that of the original subject-specific version. The methodology carried out employs a subjectspecific algorithm, sourced from a Kaggle competition, which utilizes a Support Vector Machine and spectral power and correlation coefficients as its features. The training of the classifier was modified to be subjectindependent and then compared to the performance of the subject-specific version. The results indicate that the subject-independent version performed worse than the original subject-specific one, in fact it performs below or equal to random prediction rates. It is concluded that: due to the dependency of epileptic seizure prediction algorithms on the strict characteristics of a subjects epilepsy, a subject-independent algorithm, produced with the adjustment of a subject-specific version, can’t, at this time, achieve prediction rates equal to or higher than that of the subjectspecific version.