Wielding and evaluating the removal composition of common artefacts in EEG signals for driving behaviour analysis

Noninvasive EEG signals provide neural activity information at high resolution, of which human mental status can be properly detected. However, artefacts always exist in brain oscillatory EEG signals and thus impede the accuracy and reliability of relevant analysis, especially in real-world tasks. Moreover, the use of a mature artefact identification method cannot assure impeccable artefact separation; this leads to a trade-off between removing contaminated information and losing valuable information. This study addresses this problem by investigating a simulator-based driving behaviour analysis using a car-following scenario to correlate the EEG-based mental features with behavioural responses. The study develops an architecture for an artefact composition pool and proposes three integrated prediction models to evaluate the removal compositions of the EEG artefacts. Three errors (mean absolute, root mean square, mean absolute percentage) and R-squared index are considered for measuring the performance of the models. The results show that the best-performing composition outperformed the no-removal and all-removal cases by 11.75% and 4.28% improvements, respectively. Specifically, we investigate different common artefacts including eye blinks, horizontal eye movements, vertical eye movements, generic discontinuities and muscle artefacts. The gained knowledge on artefact removal, EEG spectral features and stimuli-response patterns can be further applied to properly manipulate real-world EEG signals and develop an effective brain-computer interface.