Robust estimation of 1/f activity improves oscillatory burst detection.

Neural oscillations often occur as transient bursts with variable amplitude and frequency dynamics. Quantifying these effects is important for understanding brain-behaviour relationships, especially in continuous datasets. To robustly measure bursts, rhythmical periods of oscillatory activity must be separated from arrhythmical background 1/f activity, which is ubiquitous in electrophysiological recordings. The Better OSCillation (BOSC) framework achieves this by defining a power threshold above the estimated background 1/f activity, combined with a duration threshold. Here we introduce a modification to this approach called fBOSC, which uses a spectral parametrisation tool to accurately model background 1/f activity in neural data. fBOSC (which is openly available as a MATLAB toolbox) is robust to power spectra with oscillatory peaks and can also model non-linear spectra. Through a series of simulations, we show that fBOSC more accurately models the 1/f power spectrum compared with existing methods. fBOSC was especially beneficial where power spectra contained a 'knee' below ~.5-10 Hz, which is typical in neural data. We also found that, unlike other methods, fBOSC was unaffected by oscillatory peaks in the neural power spectrum. Moreover, by robustly modelling background 1/f activity, the sensitivity for detecting oscillatory bursts was standardised across frequencies (e.g., theta- and alpha-bands). Finally, using openly available resting state magnetoencephalography and intracranial electrophysiology datasets, we demonstrate the application of fBOSC for oscillatory burst detection in the theta-band. These simulations and empirical analyses highlight the value of fBOSC in detecting oscillatory bursts, including in datasets that are long and continuous with no distinct experimental trials.