Abstract
Blind source separation (BSS)-based hybrid algorithms are useful for separating distinct neural sources or artifacts that are mixed in electroencephalography (EEG). However, the recovered components are typically unlabeled and lack clear associations with specific underlying neural processes or artifacts. This study developed and validated a novel algorithm called EEGLabeler that uses machine learning techniques to automatically assign meaningful labels (e.g., horizontal ocular, vertical ocular, P300, and other components) to BSS-recovered EEG components based on their unique spatial characteristics. The results demonstrated that EEGLabeler achieved good performance, with macro accuracy, sensitivity, specificity, precision, and F1 scores of 0.96, 0.93, 0.96, 0.77 and 0.83, respectively, using Neural Network classifier. When using k-Nearest Neighbor (kNN) classifier, the macro accuracy, sensitivity, specificity, precision, and F1 scores are 0.96, 0.83, 0.93, 0.80, and 0.81, respectively. Both models demonstrated good cross-individual consistency and robustness. Additionally, we optimized the EEGLabeler algorithm by incorporating data augmentation and label imbalance handling as pre-processing methods. While data augmentation did not significantly improve performance, label imbalance handling enhanced the algorithm performance (increasing macro F1 from 0.81 to 0.83 for the kNN classifier, and from 0.77 to 0.79 for the Neural Network classifier). This study highlights the practicality and viability of combining BSS with machine learning models for automated identification of neural and artifact components from EEG. It opens up new possibilities for real-time EEG applications, e.g., source-based brain-computer interfaces, where the automatic extraction of neural signals from artifacts is critical for effective communication between the brain and external devices.
Authors
Chan, G. H., Hsiao, J. H. W., Gao, J., Wong, A. Y. L., Fong, G. C., Privitera, A. J., ... & Sun, E. R. Eeglabeler
False