A 2023 double-blind, sham-controlled study by Uslu and Vögele explored a deceptively simple question: in neurofeedback, is more control better—specifically, does learning faster and choosing your own pacing translate into sharper cognition? This paper is recent, but is old enough to be considered from the archives; its ideas remain current in a field that is increasingly interested in personalization, engagement, and “why it works for some people and not others.”

Neurofeedback (and its close cousin, biofeedback) are training approaches that use real-time physiological signals as a mirror: your nervous system produces measurable activity, that activity is translated into feedback, and you gradually learn—through practice and reinforcement—to shift the signal in a desired direction. In EEG neurofeedback, the signal is brain electrical activity, often summarized as power in specific frequency bands. The promise is not just changing a number on a screen, but nudging the underlying brain state in a way that can support attention, flexibility, memory, or emotional steadiness.

The catch is that neurofeedback learning is uneven. Some people improve the target signal quickly; others stall, plateau, or drift. This paper zooms in on two practical levers clinicians and researchers argue about constantly: (1) learning rate (a measurable slope of improvement during training), and (2) self-pacing (giving the trainee control over when to pause and rest). Together, these levers speak to a bigger clinical question: should we treat neurofeedback like a standardized workout plan—or like skill learning, where autonomy and timing shape what the brain actually encodes?

Methods

Design and participants

The study used a randomized, double-blind, sham-controlled design with 60 healthy adults (18–35 years; predominantly female). Participants were first assigned on a 2:1 basis to either real IUA neurofeedback or sham neurofeedback, and then split 1:1 into either self-paced or externally paced training.

Cognitive testing

Participants completed two cognitive assessments before and after the neurofeedback session:

• Mental rotation task (computerized): 96 trials with 3D object pairs presented at different rotation angles (0°, 50°, 100°, 150°). Outcomes included reaction time (analyzed after log transformation) and error rate.

• Trail Making Test (paper-pencil): Part A (visual search and motor speed) and Part B (set shifting / mental flexibility). Completion times were analyzed after log transformation.

EEG acquisition

EEG was recorded with 32 Ag/AgCl electrodes placed according to the 10/20 system, referenced to FCz, with additional electrodes for horizontal eye movements. Signals were sampled at 1000 Hz. (You might wonder about the positioning of the reference electrode... since we're looking at posterior rhythms, it's a little unusual, but not necessarily inappropriate. Still, I wouldn't have done it this way.)

Neurofeedback protocol

A 5-minute eyes-open baseline resting-state recording was used to determine each participant’s IUA target based on their individual alpha peak (defined operationally as the peak alpha frequency with a +2 Hz offset for the individualized upper-alpha target range).

Training targeted IUA power averaged across occipito-parietal sites P3, Pz, P4, O1, and O2. Feedback was displayed as a bar that participants were instructed to keep above a threshold line corresponding to baseline IUA power. Artifacts were handled online by pausing feedback and displaying “NOISE” whenever signal thresholds were exceeded.

Signal processing for feedback used a sliding FFT on 1-second windows with 75% overlap (4 Hz update rate), zero-padding for 0.5 Hz frequency resolution, and Savitzky–Golay smoothing to facilitate alpha peak identification.

Sham condition and blinding

Sham neurofeedback consisted of pre-recorded feedback taken from another participant not otherwise involved in the study. The technician was visually separated from participants and could not see the feedback stream to maintain blinding.

Self-paced vs externally paced training

Total session structure was held constant: 30 minutes of training and 4 minutes of rest (34 minutes total). Only the distribution of rest differed.

• Externally paced: five training blocks separated by fixed 1-minute rest epochs.

• Self-paced: participants decided when to rest by pressing a key; rest began immediately and training resumed when they pressed again, with a short delay before feedback restarted.

Learning rate

Learning rate was calculated per participant as the slope of relative IUA power change across training epochs. The authors then tested whether learning rate predicted post-training changes in resting-state IUA and cognitive performance.

Results

Training-related IUA changes

Across time, relative IUA power increased during training in both real and sham conditions, but the increase was larger in real IUA neurofeedback. The estimated improvement over the full training period was approximately twice as large for real neurofeedback (about a 22% increase across epochs) compared to sham (about an 11% increase). This supports that the feedback contingency mattered—even when practice and expectancy effects were controlled.

Resting-state activity

At the group level, there were no clear differences in pre- vs post-training resting-state relative IUA between conditions. However, learning rate changed the story: participants who increased IUA more quickly during training also showed larger increases in post-training resting-state IUA, and this relationship was stronger (and directionally “healthier”) in the real neurofeedback group than in sham.

Mental rotation performance

Everyone improved somewhat from pre to post, consistent with practice effects. The headline finding was not a simple “real beats sham” result. Instead, learning rate differentiated outcomes: in the real neurofeedback condition, faster increases in IUA during training were associated with fewer mental rotation errors after training. The learning-rate relationship was meaningfully different in the sham group, suggesting that “getting better at the bar” is not the same thing when the signal is not actually yours.

Reaction time decreased from pre to post across groups, but the relationship between learning rate and reaction time was more nuanced and did not cleanly support a straightforward benefit in the real neurofeedback group.

Trail Making Test

Part A (speed/visual search) improved for everyone (practice effect). In Part B (higher-level flexibility), self-paced participants improved more than externally paced participants. Notably, this self-pacing advantage appeared regardless of whether feedback was real or sham, suggesting that pacing itself—autonomy over rest and task rhythm—can support executive performance, even when the neurofeedback contingency is not the active ingredient.

Discussion

This paper offers a useful clinical reality check: neurofeedback can successfully shift the target signal during training, but the meaning of that shift depends on how the change unfolds within the individual. The most instructive effects were not simply “real versus sham,” but the way learning rate predicted outcomes specifically when the feedback was truly contingent on the participant’s own EEG.

One practical takeaway is that learning rate may be one of the most actionable variables we can track in sessions. In day-to-day practice, we often notice that two people can finish the same number of runs with similar-looking averages, yet one person’s curve shows a steady upward drift while another is jagged, flat, or only improves late. This study treats that curve as data rather than vibes: the slope of improvement was linked to post-training resting-state IUA shifts and to reduced errors in a spatial cognition task. In other words, the session may be teaching the brain something that carries beyond the training screen, but primarily when the brain is actually learning (not merely performing).

The self-pacing findings add another layer. Autonomy over rest breaks improved Part B trail-making performance—one of those tasks that tends to punish cognitive rigidity and reward flexible shifting. A plausible mechanism is that self-pacing lets participants manage arousal and fatigue more intelligently than an externally imposed schedule. Rest is not just “dead time”; it is part of the learning loop. In skill learning, the brain consolidates patterns in micro-pauses. Neurofeedback may follow a similar logic: the nervous system needs moments to recalibrate, test strategies, and return to the target with renewed signal quality.

The lack of a strong group-level cognitive advantage for real over sham in a single session is also important. It suggests that “one-and-done” cognitive enhancement claims should be treated cautiously, especially when the control condition is rigorous. A more realistic view is that neurofeedback changes may accumulate across sessions, and that measurable cognitive benefits might depend on (a) achieving a meaningful learning rate, (b) targeting a signal that truly maps onto the cognitive function being tested, and (c) designing the training in a way that supports engagement and reduces artifact-driven frustration.

Finally, the authors point to a limitation that clinicians will recognize immediately: training one narrow band may not be the whole story. If nearby bands or broader state variables (like vigilance or fatigue) shift during training, they could drive performance changes more than the nominal target. Practically, this argues for careful assessment, thoughtful protocol selection, and ongoing measurement rather than assuming that increasing upper alpha is universally the “correct” direction for everyone.

If you’d like, I can add a short interpretive paragraph here that links these results to broader learning theory (operant conditioning versus implicit learning, autonomy and reward sensitivity, and how we might measure “responders” without turning them into a binary label). I can also include a small set of supporting citations (with full APA references + links) beyond the primary paper.

Brendan’s perspective

1. Learning rate as a clinical vital sign

In this study, the most clinically useful variable was not simply whether IUA went up, but how it went up. A learning rate is basically the slope of skill acquisition: does the brain show a steady upward drift in the target signal across runs, or does it wobble, stall, and only spike when the stars align? In clinic, that slope becomes a vital sign because it tells me whether we are actually training self-regulation, or just watching noise.

Practically, I treat the learning curve like a dashboard. If a client’s IUA curve is trending upward early, I am more comfortable maintaining the same target and gently tightening thresholds over time. If it is flat or chaotic, I do not automatically “push harder.” I slow down and ask a different set of questions: is the state we are asking for even reachable right now (sleep, caffeine, stress load, pain)? Are artifacts stealing reinforcement? Is the feedback too delayed or too punishing? Is the person stuck in an over-effort strategy that spikes beta and collapses alpha?

When learning rate is low, I will often reshape the session before I change the protocol. That can mean shorter runs (to reduce fatigue), more frequent micro-breaks, coaching a lighter attentional stance, or pairing in a quick primer like paced breathing or heart rate variability biofeedback to stabilize arousal. If the curve remains stubborn, then I start thinking about protocol individualization: shifting sites, changing the reward band, adding an inhibit for high beta or slow drift, or moving to a different target altogether (for example SMR at Cz, 12–15 Hz, when the system needs more stability before it can handle upper alpha work). The key is that learning rate turns “non-responder” from a label into a solvable clinical puzzle.

This is obviously an oversimplification. It's really hard to express how complex it is to interpret data that is often highly dynamic, is multifactorial and subject to mediation and moderation effects, and finally also needs to consider subjective experience (both on the part of the trainer and the trainee).

2. Upper alpha in the real world

Upper alpha is a bit like the brain’s “quietly ready” gear: not sleepy, not tense, but organized enough for efficient processing. The study’s approach is a clean example of how I like to think about alpha work: define each person’s individual alpha peak and then train an individualized upper-alpha band rather than forcing everyone into a generic 10–12 Hz box.

In practice, when I train individualized upper alpha, I am usually targeting occipito-parietal sites (often Pz with flanking P3/P4, and sometimes O1/O2 depending on the presentation and signal quality). The goal is not to make someone feel like they are meditating harder; it is to reinforce a state where visual–spatial processing and attentional allocation become more efficient. The coaching is paradoxically simple: less striving, more “clear and settled.” If a client pursues alpha like a performance metric, they often generate compensatory tension that pushes the EEG in the wrong direction.

Upper alpha is not a universal good. In clients who already struggle with low energy, dissociation, or excessive daydreaming, aggressive alpha uptraining can backfire by nudging them further away from task engagement. In those cases, I might prioritize stabilization first (SMR 12–15 Hz at Cz, or low beta 15–18 Hz at C3/C4 depending on the phenotype), and then revisit alpha later as a secondary layer.

I also rarely treat upper alpha as a stand-alone “forever protocol.” In anxiety presentations, upper alpha can be a beautiful bridge into calmer cognitive flexibility, especially when paired with arousal regulation (breath/HRV) and good sleep scaffolding. In peak performance work, it often becomes one ingredient in a larger recipe: a few minutes of IUA to reduce mental noise, then task-relevant training (for example SMR or focused beta1) to lock in execution.

3. Translating this to ADHD, anxiety, and peak performance

What I love about this paper is that it quietly argues for two forms of personalization that matter in the real world: what you train and how you structure the learning environment. The learning-rate findings tell us that neurofeedback effects are more likely to transfer when the person is truly learning self-regulation. The pacing findings suggest that autonomy over rest can support executive performance, even when everything else is held constant.

For ADHD, this is practically a blueprint. Many clients do better with shorter bouts, immediate reinforcement, and the option to reset before frustration snowballs into artifacts and disengagement. I often structure sessions with brief, consistent runs and allow flexible pauses, while keeping the overall dose stable. Protocol-wise, this might look like SMR at Cz (12–15 Hz) for behavioral inhibition and motor quieting, or theta/beta-style attention training at midline or frontal sites depending on the assessment. Learning rate becomes the guide: if the slope is improving, we stay the course; if it is flat, we adjust the learning conditions first.

For anxiety, self-pacing can be surprisingly therapeutic because it restores a sense of agency. Clients learn that they can step out of overload and return to training without “failing.” Pairing EEG work with a body-based regulator (paced breathing, HRV, or skin conductance training) often accelerates learning by reducing sympathetic spikes that disrupt signal quality.

For peak performance, pacing is often the difference between training a fragile state and training a reliable one. Athletes, executives, and artists benefit from learning how to enter the target state, hold it, and then deliberately recover. In that context, a self-paced structure is not indulgent; it is rehearsal for real-world performance cycles.

Sham = placebo? (Hint: NO. Not even close.)

A final note, because it sits in the background of this entire discussion: sham conditions are not “nothing.” They can improve performance through expectancy, task practice, engagement, and the very human act of trying to master a challenge. I have a peer-reviewed article currently in press that will unpack this idea more fully. In the meantime, it is worth revisiting our NeuroBLOG post on the Ninaus (2013) paper on control beliefs and neurofeedback learning, because it pairs beautifully with the present study’s message: the brain learns best when the person believes their actions can meaningfully shape the outcome.

Conclusion

This study reinforces two ideas that deserve more attention in neurofeedback conversations: the curve matters, and control matters. Real IUA neurofeedback produced stronger training-related increases in IUA than sham feedback, but the more clinically meaningful signal was the learning rate—how consistently and quickly participants improved during the session. That learning rate predicted increases in post-training resting-state IUA and fewer errors on a demanding spatial cognition task.

Self-pacing added a separate benefit: participants who could decide when to rest showed greater improvement on a measure of mental flexibility. Even without a strong “real versus sham” cognitive advantage after a single session, the findings suggest that neurofeedback outcomes are not only about the protocol, but also about how the training is experienced. When we treat neurofeedback as skill learning—measuring learning curves and designing autonomy into the session—we may give the brain the conditions it needs to do what it does best: adapt.

References

Uslu, S., & Vögele, C. (2023). The more, the better? Learning rate and self-pacing in neurofeedback enhance cognitive performance in healthy adults. Frontiers in Human Neuroscience, 17, 1077039. https://doi.org/10.3389/fnhum.2023.1077039