EEGOpt: A performance efficient Bayesian optimization framework for automated EEG signal classification.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Accurate electroencephalography (EEG) signal classification requires optimal combinations of signal processing, feature extraction, and classification methods.
• Identifying the best methods for diverse EEG applications presents a significant challenge due to the lack of a universal approach.

Purpose of the Study:

• Propose EEGOpt, a Bayesian optimization framework to automate and optimize methodological choices in EEG signal processing and classification.
• Enhance the accuracy and efficiency of EEG data analysis for various applications.

Main Methods:

• Utilized Tree-Structured Parzen Estimator (TPE) for optimizing denoising (Empirical Mode Decomposition, Wavelet Packet Decomposition), feature extraction (spatiotemporal, nonlinear, spectral), and classifier selection.
• Implemented a modular caching mechanism to reduce redundant computations during the optimization process.
• Evaluated EEGOpt on three datasets, benchmarking against deep learning models (EEGNet, ShallowConvNet, DeepConvNet) and comparing TPE with other sampling methods.

Main Results:

• Achieved a maximum classification accuracy of 99.63%, significantly outperforming deep learning models (EEGNet: 96.20%, ShallowConvNet: 90.83%, DeepConvNet: 90.29%).
• The caching mechanism reduced computation time by 74.69% (vs. no caching) and 95% (vs. deep learning models).
• Identified optimal parameters for music-based EEG classification: covariance and wavelet features, k-nearest neighbor classifier, and Wavelet Packet Decomposition (WPD) denoising.

Conclusions:

• EEGOpt provides a scalable, interpretable framework for automated EEG analysis.
• The framework adapts signal processing and classification strategies to specific EEG datasets.
• EEGOpt is a valuable tool for advancing neuroscientific research, diagnostics, and brain-computer interface development.