A guide towards optimal detection of transient oscillatory bursts with unknown parameters.

Recent event-based analyses of transient, intermittent burst-like activities in the beta and gamma frequency ranges have characterized the oscillatory bursts as a neural signature that bridges dynamic neural states to cognition and behavior. While a precise detection of burst events is crucial for inferring their relations to behaviors, large variations of the background noise in the signal poses challenges for precisely identifying their onsets. Here, we examined several classic burst detection algorithms and their robustness to noise by comparing their ability to extract bursts under different conditions of signal-to-noise ratio and event duration using synthesized signals containing bursts of multiple frequencies. Our findings revealed that the detection of bursts is heavily influenced by event duration, while the precise identification of burst onsets is relatively more susceptible to the noise level. Given that burst properties in real signals are typically unknown in advance, we proposed a selection rule that utilizes the empirical cumulative distribution function and its associated area under the curve as potential criteria for determining the most suitable algorithm for a given dataset. To validate our rule, we applied the selected algorithm to theta, beta, and gamma activities in the basolateral amygdala of male mice during exposure to a spider robot-induced threat. The chosen method exhibited high detection and temporal accuracy, although statistical significance was not consistently observed between the different algorithms across frequency bands. Notably, the algorithm selected by human visual screenings did not always align with the algorithm recommended by our selection rule, indicating a mismatch between human priors and mathematical assumptions embedded in the algorithms. Consequently, our proposed rule offers a potential solution for algorithm selection in burst detection; however, its implementation also exposes the limitations of these algorithms, highlighting their variable performance depending on the dataset and cautioning against relying solely on heuristic-based approaches.