EEG Signal Classification with Data Augmentation for Epileptic Focus Localization and Deep Sleep Detection.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is vital for neurodiagnostics like epileptic focus localization and sleep stage detection.
• Limited annotated EEG data restricts deep learning model performance and generalization.
• Deep learning models require diverse and robust training datasets for clinical applications.

Purpose of the Study:

• To propose a unified EEG classification framework utilizing lightweight data augmentation.
• To enhance the robustness and generalization of deep learning models for EEG analysis.
• To investigate the impact of simple augmentation techniques on classification performance.

Main Methods:

• Implemented three lightweight data augmentation techniques: time shifting, amplitude scaling, and noise addition.
• Evaluated the framework using DeepConvNet, ShallowConvNet, and EEGNet.
• Tested the models on two public EEG datasets for physiological and pathological tasks.

Main Results:

• Data augmentation consistently improved classification performance across all tested models and tasks.
• Significant performance gains were observed: up to 2.06% for deep sleep detection and 4.07% for epileptic focus localization.
• Augmentation provided additional benefits even when baseline model accuracy was already high.

Conclusions:

• Lightweight data augmentation effectively enhances the robustness and classification performance of EEG-based deep learning models.
• Simple augmentation strategies are particularly beneficial under data-limited conditions.
• The proposed framework offers a practical approach to improving EEG analysis accuracy in clinical neurodiagnostics.