ISSN: 1223-1533

< Return to the list of articles

COMPARISON OF METHODS FOR BURST DETECTION IN NEURAL ACTIVITY


Authors: Gorana Mijatovic, Tatjana Loncar-Turukalo, Emmanuel Procyk, Dragana Bajic




 

 

 

Figure 41.1. Comparison between RS and PS methods

 

 

Introduction: It is considered that bursting neural activity provides reliable and selective neural communication, allowing for dynamical brain reorganization. Despite the importance, bursting activity lacks clear definition, being largely dependent on burst detection methods. In this study performance of parametric Poisson surprise (PS) burst detection method is compared to non-parametric rank surprise (RS) method. RS method relies on generally accepted definition of burst as occurrence of many spikes in small time interval.

 

Methods: Analysis was done on neural ensemble of 56 inter-spike-interval (ISI) series of behaving macaque monkeys. PS method compares number of spikes in certain time period to number that would be observed in Poisson process with the same average firing rate (FR). Burst extraction using PS statistic is done using maximization of the surprise statistic (SM). For rank surprise statistic the Poisson assumption is avoided and rank of each ISI is considering. The burst extraction searches for the maximum of RS statistic using dubbed exhaustive surprise maximization (ESM).

 

Results: Performance of burst detection algorithms is evaluated on ISI series comparing the percentage of spikes involved in a burst, number of bursts and burst’s length. Combining SM detection with PS and RS statistics reveals that RS is more robust to the change in ISI distribution. The comparison of SM detection with PS and ESM with RS revealed that number of spikes within bursts is in average 5,33% larger for RS then for PS under same criteria. It is determined that in average burst length is 10,15 % larger and number of bursts 6,58 % larger for RS statistics.

 

Conclusion: Bursting activity can be well identified using RS statistic as an alternative to PS statistics. RS statistics makes no assumptions on ISI distribution and performs robustly on non-Poisson firing patterns based on a very general definition of a burst.