Eye state classification based on electroencephalogram decomposition considering the dipole nature of brain signals

Abstract

Classification of brain activity states from electroencephalography (EEG) data is a relevant task for neuroscience and brain-computer interfaces. In previous works, spectral analysis of electrode signals provided eye state classification accuracy of 70–80%, while quantitative recurrence analysis with an SVM classifier achieved up to 95% for occipital electrodes, but both approaches operated on mixed signals from multiple sources and artifacts, limiting accuracy and physiological interpretability.

The aim is to develop an approach to eye state classification (open/closed) based on independent component analysis (ICA) of multichannel EEG data with component validation through equivalent dipole fitting.

The EEG Motor Movement/Imagery Dataset (109 participants, 64-channel EEG) was used. A sequential processing pipeline was implemented in Julia using the NeuroAnalyzer.jl toolbox: bad channel detection (flat, amp, var methods), frequency filtering (highpass 2 Hz, notch 50 Hz), ICA decomposition, and equivalent dipole fitting. An original loss function was developed with L2 regularization and ellipsoidal anatomical constraint based on a three-dimensional brain model.

Dataset quality analysis showed that only 79 of 109 participants (72%) had sufficient reliable channels and components with dipole scores above 85%. Among ICA components, sources with characteristic alpha, mu, beta, and gamma rhythms were identified, confirming physiological interpretability of the method.

Neural network classification based on 507 features (dipole score, variance, spatial dipole characteristics, power spectral density of three most powerful components) achieved 97% accuracy on training and 90% on test data. An overfitting problem was observed due to the small dataset size (158 samples), requiring the application of early stopping.

Unlike previous works, the proposed approach operates on separated brain sources, providing greater physiological interpretability and opening perspectives for mathematical modeling of cleaned ICA signals using systems of differential equations.