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

Accurate classification of electroencephalography (EEG) signals depends on the optimal combination of signal processing, feature extraction, and classification methods. Since no single approach is suitable across different domains, identifying the best methods for each application remains a critical challenge.

We propose EEGOpt, a Bayesian optimization framework designed to automate and optimize methodological choices in electroencephalography (EEG) signal processing and classification.

EEGOpt employed Tree-Structured Parzen Estimator (TPE) to optimize signal denoising, feature extraction, and classifier selection. The search space included Empirical Mode Decomposition and Wavelet Packet Decomposition (WPD) for denoising; spatiotemporal, nonlinear, and spectral features; and classifiers with distinct decision boundaries. A modular caching mechanism was used to minimize redundant computations. EEGOpt was evaluated on three datasets and benchmarked against deep-learning models (EEGNet, ShallowConvNet, and DeepConvNet). TPE was compared with sampling methods, including Gaussian Process, Covariance Matrix Adaptation Evolution Strategy, Quasi-Monte Carlo, and random search.

EEGOpt achieved classification accuracies of up to 99.63 %, outperforming EEGNet (96.20 %), ShallowConvNet (90.83 %), and DeepConvNet (90.29 %). The caching mechanism reduced computation time by 74.69 % compared to no caching, and by 95 % compared to deep learning models. TPE was effective in navigating hierarchical search spaces to locate global optima. EEGOpt identified covariance and wavelet features, k-nearest neighbor classifier, and WPD denoising as optimal for music-based EEG classification.

EEGOpt is a scalable and interpretable framework that automatically identifies optimal signal processing and classification strategies adaptable to EEG datasets, making it a valuable tool for neuroscientific research, diagnostics, and brain-computer interface development.