Abstract
This study assessed whether resting-state quantitative EEG (qEEG) can differentiate tinnitus laterality under rigorous multiple-comparison control and nested, cross-validated machine learning (ML). We analyzed 210 pre-specified qEEG features-spectral power (n = 95), functional connectivity (n = 80), and hemispheric asymmetry indices (n = 35)-in 110 patients with chronic tinnitus (bilateral = 58, left = 27, right = 16, non-lateralized = 9). Group differences were tested with Kruskal-Wallis tests and Benjamini-Hochberg false discovery rate (FDR) correction (q < 0.05), followed by exploratory pairwise Mann-Whitney U tests and ANCOVA controlling for left-ear pure-tone average (PTA_L). No feature survived FDR in the four-group comparison. Although 48 features showed uncorrected p < 0.05 in at least one pairwise contrast, none remained significant after FDR. In the left-right contrast, effect sizes were small and post hoc power was low (< 30%). ANCOVA indicated that most apparent differences were attributable to hearing asymmetry: after adjustment, no features survived FDR, covariate effects were small (median partial η(2)≈0.01), and results were unchanged after adjusting for interaural PTA asymmetry or mean PTA. Power modeling indicated that approximately 335 participants per group would be required to detect effects of d≈0.23 with 80% power. Nested ML models (random forest, SVM, logistic regression) performed at chance in four-class classification and near-chance in binary contrasts (balanced accuracy 57-63%; ROC AUC≈0.56), mirroring weak univariate effects. Overall, resting-state scalp qEEG showed no robust cortical biomarkers of tinnitus laterality after multiple-comparison correction and adjustment for hearing thresholds; larger, balanced cohorts will be essential for future biomarker discovery.