Frequency-Specific Operant Learning in Neurofeedback Reveals Distinct Cortical Mechanisms: Evidence from Double-Blind ERSP and ERP Dissociations

Abstract

Background: Neurofeedback reliably alters EEG activity, but two questions remain unresolved: what cortical mechanism underlies reward-contingent learning, and whether training produces durable resting-state change rather than transient within-session shifts. No study has examined reward-locked event-related spectral perturbations (ERSP) under double-blind, active-placebo-controlled conditions. Methods: Forty participants underwent five training sessions of single-channel EEG biofeedback (C3 SMR 12-15 Hz, n = 8; C3 Beta 15-18 Hz, n = 8; C4 SMR 12-15 Hz, n = 8; active-placebo sham, n = 16), plus a retention session a median of 29 days post-training, with concurrent 64-channel EEG recording. ERSP was computed from reward-locked epochs (approximately 600-700 trials per session) using Morlet wavelets (3-40 Hz) across four sessions. Resting-state alpha trajectories were modeled with growth-curve mixed-effects models on the pipeline's fixed 8-12 Hz alpha band, with an IAF-anchored reanalysis (IAF +/- 2 Hz) as a robustness check. Results: Under double-blind active-placebo control, SMR (but not Beta) neurofeedback produced a durable elevation of resting eyes-closed alpha persisting to the one-month follow-up (median 29 days). An LME growth curve confirmed SMR-specific accumulation (C3 SMR x Session B = 1.44, p = 0.004; C4 SMR B = 1.24, p = 0.012; Beta and sham flat), robust to IAF-anchored reanalysis (C3 SMR B = 1.21, p = 0.006; C4 SMR B = 1.10, p = 0.013) and present in 13 of 16 (81%) SMR participants versus 7 of 16 sham. Critically, the Beta group showed the largest immediate reward-locked ERD (d = -2.38 vs sham) yet the weakest consolidation, dissociating acute control from durable plasticity. Active groups produced frequency-specific event-related desynchronization (ERD) in the rewarded band (pooled Active vs Sham d = -1.23, p_adj = 0.001; C3 Beta and C4 SMR FDR-significant, |d| >= 1.12; C3 SMR trended, d = -0.80, p_adj = 0.081), absent in sham. A double dissociation emerged at C3: beta training produced the strongest ERD (d = -2.38), whereas C3 SMR training produced the largest P2 suppression (d = -1.33, BF01 = 0.10; smaller P2 at the trained electrode), consistent with distinct frequency-specific operant signatures. ERD magnitude did not track within-session resting shifts (r = -0.09, p = 0.67) and showed no detectable association with durable change at the individual level, consistent with consolidation as a property distinct from acute control. An ICA-based sensitivity analysis confirmed convergence of all primary findings. Conclusions: Neurofeedback engages frequency-specific, contingency-dependent cortical mechanisms whose consolidation profiles differ by protocol. The Beta arm controls the rhythm most strongly during training yet does not consolidate; both SMR arms show weaker acute control but durable resting-state growth. Consolidability, not session count or acute control, is the property that distinguishes protocols. The ERD-P2 dissociation suggests that beta and SMR training engage distinct frequency-specific mechanisms (C3 Beta: stronger spectral ERD with preserved P2; C3 SMR: moderate ERD with suppressed P2 at the trained site) with different capacities for offline consolidation. These findings support a multi-timescale model in which durable plasticity emerges only when the engaged circuit supports offline consolidation, a property not predicted by the magnitude of acute reward-locked control.