Approaches to characterizing oscillatory burst detection algorithms for electrophysiological recordings

Cognitive processes are associated with fast oscillations of the local field potential and electroencephalogram. There is a growing interest in targeting them because these are disrupted by aging and disease. This has proven challenging because they often occur as short-lasting bursts. Moreover, they are obscured by broad-band aperiodic activity reflecting other neural processes. These attributes have made it exceedingly difficult to develop analytical tools for estimating the reliability of detection methods.

To address this challenge, we developed an open-source toolkit with four processing steps, that can be tailored to specific brain states and individuals. First, the power spectrum is decomposed into periodic and aperiodic components, each of whose properties are estimated. Second, the properties of the transient oscillatory bursts that contribute to the periodic component are derived and optimized to account for contamination from the aperiodic component. Third, using the burst properties and aperiodic power spectrum, surrogate neural signals are synthesized that match the observed signal's spectrotemporal properties. Lastly, oscillatory burst detection algorithms run on the surrogate signals are subjected to a receiver operating characteristic analysis, providing insight into their performance.

The characterization algorithm extracted features of oscillatory bursts across multiple frequency bands and brain regions, allowing for recording-specific evaluation of detection performance. For our dataset, the optimal detection threshold for gamma bursts was found to be lower than the one commonly used.

Existing methods characterize the power spectrum, while ours evaluates the detection of oscillatory bursts.

This pipeline facilitates the evaluation of thresholds for detection algorithms from individual recordings.