Evaluating SMOTE-ENN impact on EEG signal seizure detection and classification

Epileptic seizures are a key aspect of epilepsy which leads to disturbance in brain electrical signals. Detection and classification of seizure improve the traditional method and provide support to doctors and patients. The researchers were working towards improving epilepsy seizure detection through various techniques. A challenge post due to fact that seizure onset have short duration which results in data imbalance problem in comparison to interictal and preictal states. This was a work toward handling imbalance in data to further improves the learning techniques and processes for seizure detection and classification to increase the efficiency of healthcare system. This paper analyse impact of oversampling on EEG signal seizure detection using SMOTE with Edited Nearest Neighbour.In this work, Epileptic Seizure binary classification was performed through ten state-of-art classifiers on benchmark public Bonn dataset. The results showed improvement in prediction of minority class after SMOTE-ENN approach.After SMOTE-ENN approach improvement in recall and F1 score of minority class had been observed for SGD followed by LR,KNN and SVM. For these classifiers average improvement was of around 35% for recall and 36% for F1 score.The AUC improvement of around 8-10% shown by LR,KNN,SVM and SGD.While conducting analysis it is found that Decision Tree, Gradient Boost and XGBoost classifier outperformed LR, Support Vector Machine (SVM), NB, KNN, Random Forest, Extra Tree, SGD classifiers with 100% Accuracy and ROC = 1.000.