A Method for Automated Performance Evaluation and Tuning of Neural Oscillation Detection Algorithms

Abstract

There are numerous methods to detect and characterize transient oscillatory events found in electrical signals from the brain representing local activity. These events can be used to time electrical stimulation to modulate activity and generate a behavioral response. These methods contain many parameters that are typically selected by hand. The application of automated optimization approaches allows these tuning parameters to be set in a systematic way. Five detection and characterization methods are explored. The F score was chosen from a number of performance metric to use as an objective function for tuning. The datasets are one synthetic dataset with realistic noise spectrum and events, and one real primate dataset. For the tuning, the goal is to maximize the F score across tuning parameter space. Two optimization approaches are used in this work: (1) univariate grid search, which globally searches each axis of the tuning parameter space in succession and (2) creeping random search, which randomly explores neighborhoods of points around a starting point and moves to any better point found. Using the metric of F score each method was shown to improve from the initial conditions and hand-tuned conditions. This work suggests that to obtain improved burst detection and characterization, tuning should be performed over all possible tuning parameters and the metric to use as the objective function should be the F score thus eliminating the need to hand-tune method specific parameters.