Epileptic Signal Classification Based on Synthetic Minority Oversampling and Blending Algorithm

The scalp electroencephalogram (EEG) has been extensively studied for epileptic signal classification in the past, but little attention has been paid to the data imbalance among different epileptic states. It is well known that, in general, the duration of seizure onset is less than several minutes or even shorter. This will result in an imbalance problem when comparing to the durations of the preictal and interictal states. In this article, a novel epileptic classification and seizure detection algorithm for imbalanced data is proposed. The wavelet packet decomposition (WPD)-based statistical features (SFs) of multichannel EEGs are first extracted for representation. Then, the K-means synthetic minority oversampling technique (K-means SMOTE) is applied for data balancing. A blending algorithm that consists of random forests (RFs), extremely randomized trees (Extra-Trees), and gradient boosting decision trees (GBDTs) is finally adopted for feature learning and epileptic signal classification. The developed algorithm provides an average accuracy of 89.49% and 83.90% on the Children's Hospital Boston (CHB)-MIT and iNeuro databases, respectively. For the patient-specific classification experiment on the iNeuro database, the proposed algorithm achieves the highest average accuracy of 92.68%.

Automated summary

This article introduces a novel algorithm for epileptic classification and seizure detection for imbalanced data, utilizing statistical features extracted from multichannel EEGs and applying data balancing techniques, followed by a blending algorithm for feature learning and epileptic signal classification, achieving high accuracy rates in the experiments.