Does Data Cleaning Improve Brain State Classification?

BackgroundNeuroscientists routinely seek to identify and remove noisy or artifactual observations from their data. They do so guided by the belief that including such data would reduce power to detect effects of interest. Whereas standard methods can exclude certain well defined noise sources (e.g., line noise), most forms of noise do not easily separate from signals of interest. Here we ask how well methods routinely used to "clean" human electrophysiological recordings actually boost power to detect brain-behavior correlations. New MethodThis, to the authors knowledge, is the first large-scale study of the impact of intracranial EEG preprocessing on brain state classification. ResultsWe find that several commonly used data cleaning methods (automated methods based on statistical properties of the signal and manual methods based on expert review) reduce statistical power for both univariate and multivariate classification of successful memory encoding, a behavioral state with very well-characterized electrophysiological biomarkers. Comparison with Existing MethodsBy reallocating resources towards collecting more within-patient data instead of attempting to "clean" data, neuroscientists may see increases in the statistical power to detect physiological phenomena. ConclusionsThese findings highlight the challenge of partitioning signal and noise in the analysis of brain-behavior relations. They also prescribe increases in sample size and numbers of observations, rather than data cleaning, as the best approach to improving statistical power.