A Comparative Evaluation of Linear and Non-Linear Combat Variants for Multisite EEG Harmonization

Aggregating multisite electroencephalography (EEG) data is essential for overcoming the limitations of single-site studies and achieving robust statistical power, particularly for the identification of reliable biomarkers of brain development. However, site-specific variability arising from differences in acquisition hardware and protocols often obscures genuine biological signals. This study evaluates the efficacy of three ComBat-based harmonization frameworks—NeuroComBat, NeuroHarmonize, and OPNested-ComBatGMM—in mitigating these batch effects. Resting-state EEG data from 558 healthy participants aged between 5 and 18 years were aggregated across four independent cohorts. Following standardized preprocessing, over 2000 feature–channel combinations were extracted per participant. Harmonization performance was assessed using t-SNE visualization, Z-score variance analysis, residual site means and variance dispersion analysis, and correlation-based preservation of age-related effects. All methods substantially reduced site-related variability; however, NeuroHarmonize consistently demonstrated superior performance. By modeling nonlinear age effects through Generalized Additive Models (GAMs), NeuroHarmonize achieved the lowest residual between-site dispersion while preserving biologically meaningful age–feature relationships. These findings highlight the importance of accounting for nonlinear developmental trajectories in multisite EEG harmonization and establish NeuroHarmonize as a robust strategy for large-scale developmental neurophysiological studies.