Evaluation of spectral exponent estimation methods for electroencephalographic signal

Spectral analysis of electroencephalographic (EEG) recordings classically focuses on the evaluation of the oscillatory (periodic) components. However, the power spectrum of neural signals, like other processes generated by complex systems, also exhibits an aperiodic component expressed as a power-law behavior, which has been shown to be modulated by cognitive states. As the interest in the aperiodic component has recently increased, several estimation approaches are now available, able to derive the spectral exponent of the aperiodic component from the signal power spectral density. Here we provide a comparative evaluation of two of these methods, specifically the best-known ‘fitting oscillations & one over f’ (FOOOF) and a simpler approach, which we call the ‘three steps’ procedure, also proposed in the literature. We tested their performance on simulated and real EEG sleep recordings. We also evaluated the effect of different spectral techniques, precisely the Fourier-based non-parametric spectral analysis and the parametric spectral analysis (i.e. autoregressive – AR modelling). The simulation study reported good estimation results for all the combinations of methods, but no clear winner was identified. As for the analysis of sleep EEG data, all the approaches were able to clearly distinguish the sleep stages (wake, REM and NREM) in accordance with the literature.