Combining meta and ensemble learning to classify EEG for seizure detection.

Despite two decades of extensive research into electroencephalogram (EEG)-based automated seizure detection analysis, the persistent imbalance between seizure and non-seizure categories remains a significant challenge. This study integrated meta-sampling with an ensemble classifier to address the issue of imbalanced classification existing in seizure detection. In this framework, a meta-sampler was employed to autonomously acquire undersampling strategies from EEG data. During each iteration, the meta-sampler interacted with the external environment on a single occasion with the objective of deriving an optimal sampling strategy through this interactive learning process. It was anticipated that optimal sampling strategies would be derived through interactive learning. And then the soft Actor-Critic algorithm was employed to address the non-differentiable optimization issue associated with the meta-sampler. Consequently, this framework adaptively selected training EEG data, and learned effective cascaded integrated classifiers from unbalanced epileptic EEG data. Besides, the time domain, nonlinear and entropy-based EEG features were extracted from five frequency bands (δ, θ, α, β, and γ) and were selected by Semi-JMI to be fed into this imbalanced classification framework. The proposed detection system achieved a sensitivity of 92.58%, a specificity of 92.51%, and an accuracy of 92.52% on the scalp EEG dataset. On the intracranial EEG dataset, the average sensitivity, specificity, and accuracy were 98.56%, 98.82%, and 98.7%, respectively. The experimental comparisons demonstrated that the system outperformed other state-of-the-art methods, and showed robustness in the face of label corruption.