Learning from Non-Seizure Clusters for EEG Analytics

EEG data collected in EMU is highly imbalanced and accuracy of automatic epileptic seizure detection is naturally low. Our aim is to increase the accuracy by reducing the imbalance ratio of seizure and non-seizure classes. We hypothesis that the non-seizure class itself includes various daily brain activities and then the data points are distributed as clusters in this class. In training phase, we propose a technique to cluster the majority (non-seizure) class into k clusters. Then, we train k KNN classifiers using each of k non-seizure clusters plus seizure class. In testing phase, we classify an incoming sample using this model and the non-seizure cluster closest to the incoming sample. We employed a state-of-the-art visualization technique to illustrate clusters of majority non-seizure class in two dimensions. The results, applied to MIT EEG dataset, show that our technique provides a higher average F-Measure accuracy.