Study design and closed-loop neurofeedback protocol.

(A) Experimental procedure. Following a pain-familiarization phase, participants completed two conditions (nontraining and training) in a counterbalanced order. Each condition began with a 3-min resting-state EEG recording to establish individualized α-power thresholds, followed by a neurofeedback–pain task designed to manipulate preparatory brain states prior to nociceptive stimulation. (B) Trial structure of the neurofeedback–pain task. Each trial started with a 16-s neurofeedback phase during which visual feedback was continuously updated every 400 ms based on ongoing EEG activity. In the nontraining condition, participants passively observed the feedback display, whereas in the training condition they actively attempted to regulate brain activity to control the feedback. After a variable fixation interval (3–5 s), a laser stimulus (3.5 J) was delivered to the left hand. Participants subsequently rated pain intensity and unpleasantness on separate 0–10 numerical rating scales following a brief delay (3–5 s). (C) Closed-loop neurofeedback framework. The neurofeedback system comprised four components: EEG acquisition, real-time signal processing, visual feedback, and participant self-regulation. EEG was recorded from somatosensory electrodes (C4, CP4, CP6) contralateral to the stimulation site. In the real-feedback group, feedback was contingent on participants’ ongoing neural activity; in the sham-feedback group, feedback was generated from pre-recorded resting-state EEG segments. Every 400 ms, α-power (8–13 Hz) was estimated from the most recent 800-ms EEG window. When α-power exceeded the individualized threshold, the ball moved rightward; when it fell below threshold, the ball moved rightward and downward.

Demographic and psychological characteristics of participants

Effects of neurofeedback on somatosensory α oscillations and interhemispheric connectivity.

(A) Spectral power during the neurofeedback phase. Grand-average EEG power spectra (5–20 Hz) are shown for real (red) and sham (blue) feedback under training (solid lines) and nontraining (dashed lines) conditions. The lower panel illustrates the training effect (Δpower = training − nontraining). Scalp maps depict α-band power (8–13 Hz) changes for real and sham feedback, as well as their between-group difference (real − sham). Black circles indicate electrodes over target and nontarget somatosensory regions. (B–C) Regional α-power modulation. Training selectively increased α-power at both target and nontarget somatosensory cortices in the real-feedback group, whereas no reliable α modulation was observed under sham feedback. Accordingly, the training effect was significantly greater for real than sham feedback at both sites. Violin plots show full data distributions; box plots indicate interquartile ranges, medians, and means; dots represent individual participants. (D) Interhemispheric α-band connectivity. Training enhanced α-band functional connectivity between target and nontarget somatosensory regions only in the real-feedback group, with no significant change under sham feedback. The connectivity training effect was significantly larger for real than sham feedback. (E) Relationship between connectivity and regional α-power. Within the real-feedback group, training-related increases in interhemispheric α-band connectivity were positively correlated with α-power enhancement at the nontarget somatosensory cortex. No such association was observed in the sham-feedback group. (F) Mediation model. In the real-feedback group, enhanced interhemispheric α-band connectivity mediated the training-related increase in nontarget α-power, indicating that bilateral propagation of α modulation was supported by changes in functional coupling. Significance levels: *p < 0.05, **p < 0.01, ***p < 0.001; n.s. = not significant. Abbreviations: “+” indicates an increase following training relative to nontraining; “–” indicates a decrease.

Effects of neurofeedback on preparatory EEG microstate dynamics.

(A) Spatial configurations of EEG microstates. Five canonical microstates (A–E) were identified across both real- and sham-feedback groups, with topographies consistent with prior reports: A (right-frontal/left-posterior), B (left-frontal/right-posterior), C (anterior–posterior symmetry), D (frontocentral), and E (centroparietal). (B) Training- and feedback-related modulation of microstate transitions. The left, middle, and right panels depict the significance of the main effects of training, feedback authenticity, and their interaction on microstate transition probabilities, respectively. While training influenced several transitions, the most prominent effect was a selective reduction in transitions from microstate E to microstate D under real feedback compared with sham. (C) E → D transition probability. Training significantly reduced E → D transitions during the preparatory period only in the real-feedback group, with no reliable change observed under sham feedback. Accordingly, the training effect (training − nontraining) was significantly greater for real than sham feedback. Violin plots show full data distributions; box plots indicate interquartile ranges, medians, and means; dots represent individual participants. (D) Relationship with somatosensory α modulation. Across participants, greater training-related reductions in E → D transition probability were associated with larger increases in somatosensory α power at the target site. Dots represent individual participants (red = real feedback; blue = sham feedback). The solid line indicates the linear fit, with shaded areas denoting the 95% confidence interval. Significance levels: ***p < 0.001; ns. = not significant. Abbreviations: “+” indicates an increase following training relative to nontraining; “–” indicates a decrease.

Effects of neurofeedback on subjective pain perception and pain-evoked cortical responses.

(A–B) Pain ratings. Subjective ratings of pain intensity and unpleasantness elicited by noxious laser stimulation are shown for training and nontraining conditions under real (red) and sham (blue) feedback. Training reduced both pain intensity and unpleasantness, with authentic feedback producing a significantly greater reduction in pain intensity compared with sham feedback. (C) Laser-evoked potentials (LEPs). Upper panel: Grand-average LEP waveforms for training and nontraining conditions in the real- and sham-feedback groups, together with scalp topographies of the N2 and P2 components. Lower panel: Difference waveforms (training − nontraining) and scalp maps illustrating training-related modulation of N2 and P2 amplitudes. (D–E) N2 and P2 amplitudes. Training significantly reduced N2 and P2 amplitudes for both real and sham neurofeedback groups. The training effects on these two components (training − nontraining) were significantly greater for real than sham feedback. Violin plots show full data distributions; box plots indicate interquartile ranges, medians, and means; dots represent individual participants. (F–H) Associations between pain-evoked responses, pain ratings, and preparatory brain dynamics. Across participants, greater training-related reductions in P2 amplitude were positively associated with decreases in pain intensity and unpleasantness, as well as with reduced E → D microstate transition probability. Each dot represents an individual participant (red = real feedback; blue = sham feedback); black lines indicate linear fits, with gray shading denoting 95% confidence intervals. Significance levels: *p < 0.05; **p < 0.01; ***p < 0.001; ns. = not significant. Abbreviations: T, training; NT, nontraining; “+” indicates an increase relative to nontraining; “–” indicates a decrease.

Structural equation model linking feedback authenticity, preparatory neural modulation, and pain regulation.

The structural equation model examined how feedback authenticity shaped training-related pain regulation through a sequence of neural and behavioral mediators. Feedback authenticity was specified as the exogenous variable (0 = sham; 1 = real). All mediator variables reflect training-induced changes, defined as the difference between training and nontraining conditions. M1 denotes the training-related change in somatosensory α power (Δα-power) at the target primary somatosensory cortex. M2 represents the training-related change in EEG microstate transition probability from E to D (ΔE→D transitions), indexing reconfiguration of preparatory brain state dynamics. M3 corresponds to the training-related change in pain-evoked P2 amplitude (ΔP2), reflecting modulation of cortical responses to nociceptive input. The outcome variable (Y) is a latent pain-regulation factor derived from training-related changes in pain intensity and unpleasantness ratings (Δpain ratings). Solid arrows indicate statistically significant paths (p < 0.05), whereas dashed arrows denote nonsignificant paths. Nonstandardized regression coefficients are shown adjacent to each path. Significance levels: *p < 0.05, **p < 0.01, ***p < 0.001.