A Non-Linear Support Vector Machine Approach to Testing for Migraine with Aura Using Electroencephalography

In this paper a new migraine analysis method is proposed using EEG (electroencephalography) signals to characterize migraine patients with aura (MwA). The objective of this work is to implement a technique for characterizing and extracting significant, robust and informative features from EEG signals which are representative of the interictal migraine brain state. We extract three brain characteristics using brain network analysis of alpha phase synchronization; transient abnormality analysis using wavelet scale; and finally joint time-frequency analysis using AR modeling. Feature selection and reduction techniques were performed on the sub-features of these three mutually independent features, to combat the over-fit problem as well as to maximize the generality of the classifier. Interpretation of the reduced features resembled to previous migraine studies. Furthermore, extracted features were used as inputs to a 10-fold cross validated non-linear support vector machine (SVM) classifier. The results showed a 92.9% classification accuracy for MwA in the interictal stage from the normal control (NC) group. Findings suggest electrical features for the predisposition of migraine which can lead to possible preventative interventions in the future.