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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Automated seizure detection using electroencephalogram (EEG) data faces challenges due to imbalanced seizure and non-seizure classification.
• Existing methods struggle with the persistent imbalance between seizure and non-seizure categories, hindering accurate analysis.

Purpose of the Study:

• To develop an advanced framework integrating meta-sampling and ensemble classification for improved imbalanced classification in EEG-based seizure detection.
• To address the significant challenge of imbalanced classification in epilepsy detection using EEG data.

Main Methods:

• A meta-sampler was employed for autonomous undersampling strategy acquisition from EEG data, utilizing interactive learning and the soft Actor-Critic algorithm.
• EEG features including time domain, nonlinear, and entropy-based metrics were extracted from five frequency bands (δ, θ, α, β, γ) and selected by Semi-JMI.
• The framework adaptively selected training EEG data and learned effective cascaded integrated classifiers from unbalanced epileptic EEG data.

Main Results:

• The proposed system achieved 92.58% sensitivity, 92.51% specificity, and 92.52% accuracy on scalp EEG datasets.
• On intracranial EEG datasets, the system demonstrated average sensitivity of 98.56%, specificity of 98.82%, and accuracy of 98.7%.
• Experimental comparisons confirmed the system's superiority over state-of-the-art methods and its robustness against label corruption.

Conclusions:

• The integrated meta-sampling and ensemble classifier framework effectively addresses imbalanced classification in EEG seizure detection.
• The developed system shows high performance and robustness, outperforming existing methods in epilepsy detection.
• This approach offers a promising solution for accurate and reliable automated seizure detection systems.