EEG is better when cleaning effectively targets artifacts

Abstract

Electroencephalography (EEG) is a useful tool to measure neural activity. However, EEG data are usually contaminated with non-neural artifacts, including voltage shifts generated by eye movements and muscle activity, and other artifacts that are less easily characterizable. The confounding influence of artifacts is often addressed by decomposing data into components, subtracting probable artifactual components, then reconstructing data back into the electrode space. This approach is commonly applied using independent component analysis (ICA). Here, we demonstrate the counterintuitive finding that due to imperfect component separation, component subtraction can artificially inflate effect sizes for event-related potentials (ERPs) and connectivity measures, bias source localisation estimates, and remove neural signals. To address this issue, we developed a method that targets cleaning to the artifact periods of eye movement components and artifact frequencies of muscle components. When tested across different EEG systems and cognitive tasks, our results showed that the targeted artifact reduction method is effective in cleaning artifacts while also reducing the artificial inflation of ERP and connectivity effect sizes and minimizing source localisation biases. Our results suggest EEG pre-processing is better when targeted cleaning is applied, as this improves preservation of neural signals and mitigates effect size inflation and source localisation biases that result from approaches which subtract artifact components across the entire time-series. These improvements enhance the reliability and validity of EEG data analysis. Our method is provided in the freely available RELAX pipeline, which includes a graphical user interface for ease of use and is available as an EEGLAB plugin (https://github.com/NeilwBailey/RELAX).