Comprehensive benchmarking of EEG biomarkers for Parkinson's disease: robustness, reliability, and clinical discriminability across multiple datasets

Abstract Electroencephalography (EEG) provides a high-temporal-resolution, noninvasive window into some of the abnormalities underlying Parkinson’s disease (PD), yet reported electrophysiological biomarkers vary widely across studies due to differences in preprocessing, feature extraction, and analytical pipelines. Here we provide the first comprehensive, multi-dataset benchmarking of 16 EEG-derived biomarkers, including phase-amplitude coupling, beta burst metrics, aperiodic components, complexity measures, and beta waveform features, to evaluate their robustness, reliability, and clinical discriminability. Using three independent resting-state EEG datasets, with two preprocessing conditions evaluated for one dataset, we assessed the effects of preprocessing, window length, within-session variability, channel aggregation strategies, and clinical group differences (healthy controls vs. PD OFF-medication, healthy controls vs. PD ON-medication, and PD OFF- vs. ON-medication). Cross-dataset reproducibility of group-level effects was additionally evaluated using effect-size pattern similarity, biomarker rank concordance, effect-direction consistency, and random-effects meta-analysis. Aperiodic features, entropy and complexity measures, and fractal metrics demonstrated high methodological robustness across preprocessing methods, window lengths, and aggregation strategies, and several showed disease- and medication-related effects across datasets, although effect magnitude and statistical significance varied across datasets and analytical conditions. In contrast, oscillatory biomarkers, including beta-gamma phase-amplitude coupling, burst-derived metrics, and waveform-asymmetry features, showed substantial variability across methodological choices and lower within-session consistency. However, this methodological sensitivity should not necessarily be interpreted as lower biological relevance, as these measures may capture more transient or state-dependent physiological dynamics. Fully-denoised data improved group-level separability, for several biomarkers in Dataset 1. Together, these findings identify several biomarkers with comparatively strong methodological robustness for further biological and clinical validation and highlight methodological factors that contribute to inconsistencies in the PD literature. This benchmarking framework provides practical guidance for standardizing biomarker analysis while emphasizing that methodological robustness, biological sensitivity, and clinical utility represent distinct dimensions of biomarker performance.