Single-Channel EEG Artifact Identification with the Spectral Slope

Abstract

Electroencephalogram (EEG) signals are a valuable recording technique to diagnose neurological disorders and identify noninvasive biomarkers for clinical application, however, they are vulnerable to various artifacts. It is difficult to define exact parameters which efficiently distinguish artifacts from neural activity, and thus cleaning EEG data often relies on labor-intensive visual scoring methods. While signal processing techniques to remove artifacts exist, many state-of- the-art techniques are designed for multivariate signals, which can be challenging to implement in recording setups with few electrodes. We demonstrate how the spectral slope - a method previously used to distinguish between conscious states by linear regression of the logarithmic EEG power spectra - can also be used to identify epochs contaminated by recording artifacts in rat EEG recordings and propose this as a first pass artifact detection method. We computed the mean spectral slope for both ‘clean’ and ‘noisy’ epochs and compared the distributions among individual recordings to determine whether the decision threshold should be dynamic or fixed. We found no significant difference between the mean of these distributions and determined that a spectral slope threshold of -8μV2/Hzwas effective at identifying noisy epochs across all recordings. The accuracy of our method was evaluated against visually scored recordings and obtained an average accuracy, F1 and Cohen Kappa score of 94.2%, 86.4%, and 83%, respectively, across all epochs. Our study contributes to the automation of EEG artifact detection by presenting a straightforward initial method for identifying contaminated epochs based on the spectral slope of a single EEG channel in rodent recordings.

Absence seizures and sleep abnormalities in a rat model ofGRIN2Bneurodevelopmental disorder

“…The spectral slope, a regression-based fitting measurement which describes the rate of change of the EEG power spectra, was calculated between 1 - 48 Hz using the NumPy polyfit function. 28 Epochs that were outside of the visually selected threshold of -5 log 10 (mV 2 )/Hz for the spectral slope and 500 mV for the spectral offset were removed from further analyses.…”