No Single Best Pipeline: Multiverse Analysis of EEG Preprocessing for N-Back Working Memory Tasks

Working memory (WM) involves temporary information maintenance and manipulation, widely assessed using n-back tasks. This study systematically compared forty-three electroencephalography (EEG) preprocessing pipelines, varying in high-pass filtering, low-pass filtering, Independent Component Analysis, and re-referencing methods, to evaluate their impact on event-related potentials (ERP) during the n-back (2-back and 3-back) task across three previously published datasets with a total of 164 participants. The primary outcome was ERP data quality of the P300 measured at the Pz across preprocessing pipelines. A composite signal quality score combining N200 at Fz and P300 and P200 at Pz was evaluated across all pipelines as a secondary analysis. Data quality was evaluated across three dimensions: target trials signal quality, target-nontarget difference waveforms, and group-level signal-to-noise ratios (SNR) consistency between major depressive disorder (MDD) and health control participants. No preprocessing pipeline performed optimally across all assessment dimensions. The combination of 0.5 Hz high-pass filtering, 100 Hz low-pass filtering, and REST referencing showed strengths in SNR optimization, ranking highly for target trial quality, target-nontarget difference waveforms, and group-level analyses. However, other pipelines demonstrated advantages for amplitude preservation or measurement reliability (standardized measurement error). High-pass filtering and referencing choices substantially influenced ERP outcomes at both individual and group levels. 0.5 Hz high-pass filtering alone performed optimally for target versus non-target discrimination. This multiverse analysis demonstrates that preprocessing parameters should be tailored to specific research objectives, as methodological decisions influence EEG data quality and interpretation.