Neurofeedback

Abstract

Neurofeedback is a form of biofeedback in which a person's own brain activity is measured, converted into a sensory signal in real time, and fed back so that the person can learn to change it. A recording instrument extracts a target feature of the signal, a display renders that feature moment to moment, and reward is delivered whenever the feature moves in the trained direction. Over repeated sessions the closed loop shapes the brain response through operant learning. The method spans scalp electroencephalography, slow cortical potentials, and real-time functional magnetic resonance imaging, and it has been applied to epilepsy, attention disorders, and cognitive enhancement. Its central scientific problem is separating a genuine, neural-specific effect from the powerful expectation and reinforcement effects that accompany any feedback, which modern controlled designs are built to resolve.

Keywords: neurofeedback, electroencephalography, operant conditioning

Neurofeedback closes a loop between a brain signal and the person generating it: the signal is recorded, a target feature is extracted and shown back within a fraction of a second, and the person is rewarded for moving that feature in a chosen direction (Sitaram et al., 2017). Because the reward is contingent on the person's own neural activity, the arrangement is a case of operant conditioning applied to the brain rather than to overt behaviour, and the same reinforcement principles that shape a lever press are recruited to shape a rhythm or a regional activation.

The Operant Loop

Every neurofeedback protocol is a control loop in which the controlled variable is a feature of the brain's own activity (Sitaram et al., 2017). An instrument records the signal, a processing stage isolates the target feature — the power in a frequency band, the amplitude of a slow potential, the activation of a region — and a display converts that feature into something the trainee can see or hear. When the feature moves in the trained direction, a reward marks the moment. Because reward is delivered contingent on the neural event, the arrangement is operant conditioning with a neural response as the operant, and the shaping of that response follows the same reinforcement logic that governs any operantly conditioned behaviour (Birbaumer et al., 1990).

The learning that results is well described by a saturating curve: each rewarded session nudges the trained feature further from its baseline, with diminishing returns as it approaches a ceiling set by the person's physiology and the protocol (Enriquez-Geppert et al., 2017). This up-regulation curve is the quantitative signature of successful training and is the subject of the first demonstration and the Worked Example below.

Electroencephalographic Protocols

The oldest and most widely used neurofeedback protocols read the brain with scalp electroencephalography, whose millisecond resolution makes it well suited to training oscillatory features. The founding demonstration was itself electroencephalographic: participants learned to raise and lower the alpha rhythm under auditory feedback, the first evidence that a brain rhythm could be brought under operant control and the experiment that opened the field (Nowlis & Kamiya, 1970). Three families of protocol now dominate clinical work. The first trains the sensorimotor rhythm, a 12–15 Hz rhythm over sensorimotor cortex; rewarding its enhancement was first shown to suppress seizures in an epileptic patient and remains the foundation of neurofeedback for epilepsy (Sterman & Friar, 1972; Sterman & Egner, 2006). The second trains the theta/beta ratio, the ratio of slow theta power to faster beta power, which tends to be elevated in attention disorders; rewarding a fall in the ratio is the classic protocol for attention-deficit/hyperactivity disorder and was first reported to change behaviour in a hyperkinetic child (Lubar & Shouse, 1976; Arns et al., 2014). The third trains slow cortical potentials, second-scale shifts in cortical polarisation that index the mobilisation of neural resources; learning to produce them voluntarily has been used both for self-regulation research and for communication in paralysis (Birbaumer et al., 1990; Strehl et al., 2017).

Beyond the clinic, the same electroencephalographic protocols have been applied to cognitive and affective enhancement in healthy participants, where reviews report gains in attention and memory but stress the unevenness of the evidence (Gruzelier, 2014; Enriquez-Geppert et al., 2017). The theta/beta reward rule of the second demonstration is the concrete contingency behind this literature.

Real-Time fMRI and Decoded Feedback

Electroencephalography localises poorly, so it cannot target a deep or small brain structure. Real-time functional magnetic resonance imaging removes that limit: the blood-oxygen signal from a chosen region is computed while the person lies in the scanner and fed back within a few seconds, letting the person learn to raise or lower activation in a precisely defined target (Sulzer et al., 2013). A more radical variant, decoded neurofeedback, trains a multivoxel activity pattern rather than the mean activation of a region; in a striking demonstration, participants were rewarded for reproducing the pattern associated with a visual orientation and showed perceptual learning for that orientation without ever being told what was being trained (Shibata et al., 2011). The rapid growth of these methods has been reviewed as a distinct branch of the field with its own methodological demands (Watanabe et al., 2017).

Whatever the imaging modality, the learning that neurofeedback produces is only meaningful if it is specific to the trained signal rather than a by-product of expectation and effort — the question the third demonstration and the next section address.

Efficacy and the Specificity Problem

Neurofeedback delivers reward, engages motivation, and carries a strong expectation of benefit, so any improvement it produces has two candidate causes: a neural-specific effect of regulating the trained signal, and a non-specific effect of the training context that any credible sham would also produce (Thibault & Raz, 2017). The tool for separating them is the yoked-sham control, in which a matched group receives feedback replayed from another person's brain; because that signal is uncorrelated with the trainee's own activity, only the difference between contingent and sham training measures the neural-specific effect. When outcomes are assessed by raters blind to group, meta-analyses of neurofeedback for attention disorders find that the advantage over sham shrinks markedly, with the clearest effects on the trained electroencephalographic parameter rather than on blinded symptom ratings (Cortese et al., 2016; Micoulaud-Franchi et al., 2014).

This does not show that neurofeedback fails; it shows that its evaluation must be built to isolate specificity. Rigorously controlled trials of slow-cortical-potential training in attention disorders retain effects after controlling for unspecific factors, and the field has converged on a reporting standard — the Consensus on the Reporting and Experimental Design of clinical and cognitive-behavioural neurofeedback studies, or CRED-nf checklist — that makes the control conditions and the trained-signal evidence explicit in every report (Strehl et al., 2017; Ros et al., 2020).

Worked Example

Consider the closed-loop up-regulation curve of the first demonstration, in which a trained signal is driven from a baseline of B = 1.0 toward a ceiling of C = 2.5 along the saturating path signal(t) = B + (C − B)(1 − e^−kt), with learning rate k = 0.30 per session. After five sessions the signal reaches 1.0 + 1.5 × (1 − e^−1.5) = 1.0 + 1.5 × 0.7769 = 2.165, which is 77.7% of the achievable gain from baseline to ceiling. After ten sessions it reaches 1.0 + 1.5 × (1 − e^−3.0) = 1.0 + 1.5 × 0.9502 = 2.425, or 95.0% of the achievable gain. The increment across the first five sessions, 1.165 of the 1.5 total, is more than three times the increment across the second five, 0.260 — the diminishing return that gives the up-regulation curve its shape and that explains why most protocol effects accrue early in training (Enriquez-Geppert et al., 2017).

Discussion

Neurofeedback matters because it turns an invisible internal variable into something a person can act on, extending the reach of operant conditioning from overt behaviour to the neural events that underlie it (Sitaram et al., 2017). That reach is genuine: people can learn voluntary control of a rhythm, a slow potential, or a regional activation, and in some clinical uses — most durably in epilepsy — the trained change tracks a real outcome (Sterman & Egner, 2006). The unresolved question is not whether the brain can be trained but how much of the resulting benefit is specific to the trained signal, a question that only controlled designs and honest reporting can answer (Ros et al., 2020). Seen this way, neurofeedback is a case study in how a plausible mechanism and a real learning effect can coexist with genuine uncertainty about clinical specificity, and in how a field can build the methods needed to resolve that uncertainty rather than assume it away.

Current Directions

The most active current work is methodological rather than promissory. The CRED-nf checklist has given the field a shared standard for pre-registration, control conditions, and the reporting of learning on the trained signal, and its uptake is reshaping how new trials are designed and read (Ros et al., 2020). Real-time functional magnetic resonance imaging and decoded neurofeedback continue to expand the set of targets beyond what scalp electroencephalography can reach, raising new questions about how a trained multivoxel pattern relates to behaviour (Watanabe et al., 2017; Shibata et al., 2011). And in the largest clinical application, attention disorders, the debate has shifted from whether neurofeedback works to which components of the training carry the effect, with blinded meta-analyses driving the search for the neural-specific ingredient (Cortese et al., 2016).

Common Misconceptions

Neurofeedback lets a person read or rewrite arbitrary thoughts.

Neurofeedback trains a single, pre-selected feature of the signal — a band power, a slow potential, or one region's activation — and nothing else; it neither reads the content of thought nor grants control over untrained features (Sitaram et al., 2017). What is trained is exactly what is fed back.

Any improvement after training proves a neural-specific effect.

Because training delivers reward and raises expectation, improvement can arise from non-specific factors that a yoked-sham control would reproduce; only the contingent-minus-sham difference measures the neural-specific effect (Thibault & Raz, 2017; Cortese et al., 2016).

The EEG and functional-MRI forms are interchangeable.

Electroencephalography offers millisecond timing but poor localisation, whereas real-time functional magnetic resonance imaging reaches deep, precisely defined targets at the cost of temporal resolution; the modality determines which signals can be trained at all (Sulzer et al., 2013).

Glossary

Contingent feedback.

Feedback driven by the trainee's own brain signal in real time, so that reward depends on the trainee's own neural activity.

Decoded neurofeedback.

A functional-MRI method that trains a multivoxel activity pattern rather than the mean activation of a region.

Electroencephalography.

The recording of the brain's electrical activity from scalp electrodes, with high temporal but low spatial resolution.

Neural specificity.

The portion of a neurofeedback effect that is attributable to regulating the trained signal rather than to the training context.

Neurofeedback.

A form of biofeedback in which a person's own brain activity is fed back in real time so that the person can learn to regulate it.

Operant conditioning.

Learning in which the probability of a response changes according to the reward or punishment that follows it.

Real-time fMRI.

Functional magnetic resonance imaging analysed as it is acquired so that a region's activation can be fed back within seconds.

Sensorimotor rhythm.

A 12–15 Hz electroencephalographic rhythm over sensorimotor cortex whose enhancement is the basis of neurofeedback for epilepsy.

Slow cortical potential.

A second-scale shift in the polarisation of cortex that indexes the mobilisation of neural resources and can be brought under voluntary control.

Specificity problem.

The challenge of separating a neural-specific training effect from the expectation and reinforcement effects common to any feedback.

Theta/beta ratio.

The ratio of slow theta-band power to faster beta-band power, often elevated in attention disorders and lowered in the classic attention protocol.

Up-regulation curve.

The saturating rise of a trained signal across sessions, approaching a physiological ceiling with diminishing returns.

Up-training.

A protocol that rewards an increase in the target feature, as opposed to down-training, which rewards a decrease.

Yoked-sham control.

A control condition in which a participant receives feedback replayed from another person's brain, uncorrelated with their own activity.

Key Researchers

Martijn Arns (contemporary). Director of Research Institute Brainclinics and researcher at Utrecht University; he has led work on electroencephalographic neurofeedback and the theta/beta ratio in attention-deficit/hyperactivity disorder. ORCID

Niels Birbaumer (b. 1945). Emeritus professor at the University of Tübingen; he established the self-regulation of slow cortical potentials and pioneered brain-computer interfaces for communication in paralysis. ORCID

Joe Kamiya (d. 2021). Researcher at the University of California, San Francisco; he demonstrated in the 1960s that people can learn operant control of the electroencephalographic alpha rhythm, founding the field. In Memoriam

Tomas Ros (contemporary). Researcher at the University of Geneva; he led the CRED-nf checklist consensus on the reporting and experimental design of neurofeedback studies. ORCID

Ranganatha Sitaram (contemporary). Researcher at St. Jude Children's Research Hospital; he has advanced real-time functional-MRI neurofeedback and closed-loop brain training. ORCID

M. Barry Sterman (1935–2023). Professor at the University of California, Los Angeles; he discovered the sensorimotor rhythm and established sensorimotor-rhythm feedback for epilepsy. Faculty page

Robert T. Thibault (contemporary). Metascientist at Stanford University; he has analysed the placebo and non-specific components of neurofeedback and co-authored its reporting standards. ORCID

Frequently Asked Questions

How is neurofeedback different from ordinary biofeedback? Ordinary biofeedback trains a peripheral signal such as heart rate or muscle tension, whereas neurofeedback trains a feature of the brain's own activity fed back in real time; both share the same closed-loop, reward-based structure (Sitaram et al., 2017).

Is neurofeedback a form of operant conditioning? Yes. Because reward is delivered contingent on the trainee's own neural activity, the trained brain response is an operant, and its shaping follows standard reinforcement principles (Birbaumer et al., 1990).

What was the first clinical use of neurofeedback? Enhancing the sensorimotor rhythm was first shown to suppress seizures in an epileptic patient, and epilepsy remains the application with the most durable evidence (Sterman & Friar, 1972).

Does neurofeedback work for attention-deficit/hyperactivity disorder? Theta/beta and slow-cortical-potential protocols change the trained electroencephalographic parameter, but when symptoms are rated by blinded assessors the advantage over sham shrinks, so the specific clinical benefit remains debated (Cortese et al., 2016).

What is a yoked-sham control? It is a control condition in which a participant receives feedback replayed from another person's brain; because that signal is uncorrelated with their own, only the contingent-minus-sham difference measures the neural-specific effect (Thibault & Raz, 2017).

How does real-time fMRI neurofeedback differ from the electroencephalographic form? Real-time functional magnetic resonance imaging feeds back the activation of a deep or precisely localised region within seconds, reaching targets that scalp electroencephalography cannot resolve, at the cost of temporal resolution (Sulzer et al., 2013).

What is decoded neurofeedback? It is a functional-MRI method that rewards a multivoxel activity pattern rather than a region's mean activation; participants have shown perceptual learning for a trained pattern without being told what was being trained (Shibata et al., 2011).

What is the CRED-nf checklist? It is a consensus reporting standard for clinical and cognitive-behavioural neurofeedback studies that makes control conditions and evidence of learning on the trained signal explicit in every report (Ros et al., 2020).

References