The systematic review and meta-analysis were conducted and reported in accordance with the Cochrane Handbook for Systematic Review of Intervention guidelines (Higgins et al., 2022) and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Page et al., 2021). The experimental protocol was pre-registered with the International Prospective Register of Systematic Reviews (PROSPERO) database (registration number: CRD42022357160).

To identify all research articles that contained neurofeedback interventions with healthy subjects, a comprehensive systematic literature search of three online databases was performed: PubMed, Embase, and Web of Science. The search included studies published through August 2022, with abstraction procedures conducted from August 2022 to January 2024. The terms used to identify and broadly capture all studies related to neurofeedback, included these: neurofeedback, brainwave biofeedback, EEG feedback, electromyography feedback, electroencephalographic biofeedback, and hemoencephalography.

Studies not published in English, case-reports, reviews, commentaries, letters to the editor, meta-analyses, dissertations, and conference abstracts were excluded. The inclusion criteria for research article selection was as follows: (1) the study population included participants at least 18 years of age from a non-clinical population; (2) the study design consisted of neurofeedback interventions that included feedback of brain activity presented in near real-time to the participant; (3) the intervention made use of one of the following neuroimaging methods: functional magnetic resonance imaging (fMRI), functional near-infrared spectroscopy (fNIRS), electroencephalography (EEG), or magnetoencephalography (MEG); (4) Studies reported an outcome measure reflecting a change as a result of training as it relates to acquisition or retention values; (5) A measure of variance, or information to derive variance, was reported; (6) The study was a reviewed empirical study.

Five investigators (EG, MV, DE, SM, JW) extracted data from articles generated by the previously established literature search strategy. Combinations of two investigators screened and extracted the articles, with a third investigator resolving discrepancies for consensus. All articles identified through the keyword search from the online databases were migrated to Covidence for screening firstly at the level of their titles and abstracts, followed by a second screening using the full text of the studies. Those that met eligibility criteria at this stage (N = 39) were included in the meta-analysis, at which point meta-data and parameter details were extracted.

A fully annotated list of extracted parameters and their categorizations to be compared to change in neurofeedback control can be found in the data abstraction form in the Supplementary material. Briefly, categorical parameters extracted included: study blinding, neuroimaging device used, participant characteristics (e.g., sex), strategies used for control of neurofeedback stimulus, the presence of explicit instructions given to develop control, presence of extrinsic motivation, neurofeedback display details, neurofeedback stimulus presentation timing, details of task demands between neurofeedback training intervals, signal modulation directionality, frequencies targeted, regions of interests targeted (cortical frontal, central, posterior), presence of pre-training rehearsal, presence of functional localizers.

Continuous parameters included details such as: participant characteristics (e.g., sample size, age), duration of neurofeedback training trials, intertrial interval, trials per block, interblock interval, blocks per session, intersession interval, number of sessions, duration of sessions, total neurofeedback training duration, total time spent in an experimental setting.

Neurofeedback display categories were defined as the following: simple audio refers to feedback that uses audio alone without visual components; simple visual includes stimuli that consist only of visual elements that are minimally animated, such as bars or line graphs; simple audiovisual uses stimuli that combine both minimal audio and visual components (e.g., a bar graph paired with a tone); complex displays involve stimuli that include multiple elements working together in a dynamic and interactive way (e.g., animations, games, or dynamically changing shapes).

A random-effects model was used to pool effect sizes given considerable heterogeneity between studies. The Restricted Maximum Likelihood (REML) method was used for estimating the between study variance, as it has been reported to result in reduced bias (Veroniki et al., 2016). Knapp-Hartung adjustments (Knapp and Hartung, 2003) were used to calculate the confidence interval around the pooled effect (Harrer et al., 2021). Effects were weighted using their inverse variance to account for their precision. Heterogeneity across studies was assessed using the Q and I²statistics.

Subgroup analyses were conducted to investigate the effect of discrete moderators on the SMDs while meta-regression was used for continuous moderators. Predictors with too few cases per level (< 2) were excluded from the analysis. Continuous data with skewed distributions were log-transformed before analysis. The R package meta was used for pooled effect size, subgroup and meta-regression analyses (Balduzzi et al., 2019).

The same multi-investigator resolution protocol was used with the NIH Quality Assessment Tool for Before-After (Pre-Post) Studies with No Control Group (National Heart Lung Blood Institute, 2014) to examine the quality of the selected studies. The following quality criteria were assessed: the formalization of the research question, eligibility criteria and study population, the study participants' representativeness of clinical populations of interest, consistency in enrolled participants with the research question, sample size, clarity of intervention method, the clarity and validity of outcome measures, blinding, follow-up, statistical analysis, the frequency of measurement of outcomes, and group-level intervention details. Reviewers could select “yes,” “no,” or “cannot determine/not reported/not applicable” in response to each item in the tool. Each study was given a score ranging from 1–12 with categorical thresholds labeling them as either fair (8–10), or good (10–12) to add as an additional parameter by which to assess neurofeedback training reporting. If the studies failed to meet 5 or more quality criteria, they were reviewed by the team to determine whether they posed significant risk of bias and were rated as “poor” if so. The possibility of publication bias was assessed through visual inspection of funnel plots and by using Egger's regression test (Egger et al., 1997) to detect funnel plot asymmetry.

Search keywords across three databases returned 4,243 unique articles for screening. After applying inclusion and exclusion criteria to title, abstract, and full-text screening, a total of 39 articles were found eligible for meta-analysis (Figure 1). Among these articles a total of 55 eligible participant groups were included in the analysis as separate data points. A full table with study characteristics, meta-data, and key parameters can be found in the Supplementary material.

A full table with study characteristics and meta-data can be found in the Supplementary material.

Quality assessment through the NIH Tool for Before-After (Pre-Post) Studies with No Control Group determined that there were 11 studies that were rated to have low risk of bias (“good”) while 34 were susceptible to some bias (“fair”). After careful evaluation, none of the selected studies were considered to have significant risk of bias (“poor”), therefore no studies were excluded based on the quality assessment. Studies and their scores regarding the quality assessment criteria can be found in the Supplementary material.

Analysis of change in neural activation from first to last training sessions showed a significant effect of neurofeedback training (SMD = 0.32, 95% CI 0.19 to 0.44, p < 0.001, I² = 76.0%). Leave-one-out analysis detected one influential outlier (van Son et al., 2020: Group B: 8 Sessions of training). After removing this influential case there was a small increase in the pooled effect size (SMD = 0.34, 95% CI 0.23 to 0.44, p < 0.001, I² = 71.8%). Analysis of change in neural activation from pre-training baseline to post-training also resulted in a significant effect of training (SMD = 0.26, 95% CI 0.03 to 0.50, p = 0.03, I² = 82.0%).

Total training time influenced how much neural activity changed between the first and last training sessions—studies with longer training periods showed greater changes, following a linear relationship after log-transformation. However, no clear inflection point or plateau, so the current study was not able to specify an optimal training duration. When analyzing changes in resting state activity from before to after training, the current meta-analysis indicated that total neurofeedback training duration did not affect changes in neural modulation.

After removing one outlier study, EEG studies showed a significantly higher overall effect size compared to fMRI studies. EEG studies also showed much greater heterogeneity in their results (I² = 77%) than fMRI studies, which had no heterogeneity (I² = 0%), indicating that variability observed in fMRI results was not due to between-study variability but due to sampling error. To better understand factors contributing to this heterogeneity, the current study examined several EEG-specific factors: the frequency band targeted, the number of feedback channels used, and whether one or multiple frequency bands were trained. Of these, only the target frequency band significantly contributed to the differences between studies. Most EEG studies (85%) focused on sensory-motor response (SMR), alpha, or beta bands. Other frequency bands were targeted in one study and were therefore not included in the statistical analysis.

The current results indicate that training aimed at SMR (12–15 Hz) produced greater signal changes across training sessions than training aimed at alpha (8–12 Hz). This finding is consistent with Gong et al. (2020), which reported that while participants could learn to modulate both SMR and alpha, SMR training produced larger power increases compared to rest. In contrast, alpha training led to more pronounced changes in resting-state alpha power. These results suggest that alpha neurofeedback training may have a stronger impact on resting-state brain activity, while SMR and beta neurofeedback training effects are more noticeable during active training. This in total is further evidence that success of neurofeedback may be context-dependent. It is also important to note, however, that some studies report more generalized effects of neurofeedback training. Bagherzadeh et al. (2020) for example, has found that training alpha synchrony in the parietal cortex led to changes in gamma synchrony in the parietal cortex and changes in alpha synchrony in the visual cortex.

In contrast to our hypothesis, the current results show that incorporating a pre-training rehearsal led to a smaller effect size compared to those without rehearsal. Pre-training rehearsals are practice sessions where participants attempt to modulate their brain signals, intended to help them acclimate to the task and lab environment. However, in this meta-analysis which focused on changes from first to last session, rehearsals may have elevated activity during the first session, thereby reducing the observed change between the first and last sessions of neurofeedback training.

In contrast to the current study, Haugg et al. (2021) identified pre-training rehearsals as a significant predictor of neurofeedback success—defined as modulating neural signals in the intended direction on more than 50% of trials. Unlike the current study, which focused on EEG and fMRI in non-clinical adults, Haugg et al. focused solely on fMRI studies in clinical patients. It is possible that pre-training rehearsals facilitate fMRI training but impair EEG training due to fundamental differences between the two modalities. fMRI neurofeedback is inherently delayed by the hemodynamic response, so participants may require more time to learn the mapping between their mental strategy and the feedback signal; in that context, a rehearsal does not diminish most of the training benefit. In contrast, EEG feedback is immediate and therefore pre-training rehearsal may inflate participants' initial performance.

Neurofeedback displays in the present meta-analysis were categorized into four types: simple audio, simple visual, simple audiovisual, and complex. Simple audio feedback refers to auditory-only cues—typically tones or beeps—that vary in pitch or volume to reflect ongoing neural activity. Simple visual feedback relies exclusively on minimally animated visuals, such as bars or line graphs, to convey real-time changes in the targeted brain signal. Simple audiovisual feedback combines these approaches by pairing a basic visual representation (for instance, a bar graph) with a synchronized auditory cue, thereby engaging two sensory channels simultaneously. Finally, complex feedback involves dynamic, interactive stimuli—often including animations, gamified interfaces, or richly layered audiovisual scenes—that present neural information through multiple, integrated feedback elements.

When comparing these modalities, complex feedback yielded the largest within-training improvements in neural modulation, indicating that richer, multi-element displays facilitate more robust learning. In contrast, simple visual feedback produced the smallest effect sizes. Taken together, these findings demonstrate the importance of feedback richness: complex, multi-element displays appear most effective for facilitating neurofeedback training.

The current meta-analysis did not identify additional parameters that significantly explained variability in neurofeedback outcomes. However, this absence of significant effects may reflect a limitation in data availability rather than a true lack of difference. Several subgroup comparisons were based on a small number of studies, making it difficult to detect meaningful differences even if they exist.

One area where this limitation was especially clear is the timing of feedback delivery, in which case only two out of forty-five studies used intermittent feedback, compared to forty-three studies which used continuous feedback. In operant conditioning, reinforcement schedules can be continuous—where feedback is given immediately after each behavior—or intermittent, where feedback is delivered intermittently with a fixed or varied schedule (Skinner and Morse, 1958). In neurofeedback, continuous feedback is commonly assumed to be more effective, under the premise that real-time reflection of brain activity allows participants to adjust their behavior more precisely. However, this assumption has been challenged. For example, Hellrung et al. (2018) found that intermittent feedback produced greater signal changes than continuous feedback, arguing that real-time feedback might demand excessive cognitive resources, forcing participants to divide attention between monitoring feedback and regulating their brain activity. Intermittent feedback, by contrast, may reduce cognitive load and foster a more focused internal regulation process. The current meta-analysis was not able to include enough studies using intermittent feedback to make a meaningful comparison with studies using continuous feedback.

Feedback timing may also contribute to observed differences between EEG and fMRI-based neurofeedback. While EEG supports real-time feedback, fMRI involves intrinsic delays due to the slower hemodynamic response and image acquisition process. If intermittent feedback proves advantageous in certain contexts, these delays in fMRI might align more closely with effective reinforcement timing. The current meta-analysis detected a much lower variability in neurofeedback training effects in fMRI studies compared to EEG studies, which might partly be based on differences in feedback timing.

There are several limitations of the current meta-analysis. First, the meta-analysis focused exclusively on non-clinical populations, limiting the generalizability of these findings to clinical populations. Clinical populations often demonstrate greater room for improvement and, consequently, more pronounced neural changes as a result of neurofeedback training. For example, Haugg et al. (2021) reported that clinical populations showed significantly higher success ratios in neurofeedback interventions compared to non-clinical groups, highlighting the need for studies that include diverse participant samples.

Second, there was inconsistency in how neural change was measured across studies. While some studies reported within-training outcomes—such as differences between the first and last training sessions—others reported only pre- to post-intervention changes. These outcome types are not equivalent and may capture different aspects of the learning and consolidation process. Of the 55 included groups, 35 provided within-training data, while 20 provided only pre-post comparisons. This heterogeneity complicates the interpretation and comparison of effects across studies. Additionally, overall, very few studies reported quantifiable data useful for meta-analysis, resulting in many studies being excluded during the filtering process due to insufficient data.

Third, not all parameter categories could be fully analyzed due to incomplete or uneven reporting across the literature. Certain key variables were missing from a substantial number of studies, reducing the statistical power and limiting subgroup analyses. Even among analyzable parameters, the number of groups available for comparison was often small, reducing confidence in the generalizability of parameter-specific findings.

Fourth, evidence for publication bias was found in studies that reported change from first to last training session, indicating that the interpretation of the results should be approached with caution. The field should take account of the potential publication bias present Journals can combat this by publishing high quality research regardless of its outcome. More replication studies are also needed to counteract the possible overestimation of effect sizes.

The use of Hedges' g helped correct for bias in smaller sample sizes but it does not address all potential statistical concerns, such as violations of the assumption of normality or heterogeneity of variance across studies. Finally, some parameter implementations were so infrequently used that they had to be excluded altogether. While this approach ensured analytical reliability, it may have inadvertently excluded promising experimental techniques that are still emerging within the field. We recommend future studies to follow the CRED-NF reporting checklist to increase consistency across different studies (Ros et al., 2020).

Additionally, this meta-analysis did not account for individual baseline neurophysiological factors such as menstrual cycle phase (Bazanova et al., 2017), dominant EEG frequency patterns, psycho-emotional tension (e.g., EMG) (Goncharova et al., 2003), or ergonomic conditions like eye state (i.e., open or close) and postural support. These variables could influence both the effectiveness of neurofeedback and the quality of the recorded signals. To improve comparability and interpretability across studies, future research should systematically record and report such factors.

Finally, the current study was not able to perform analyses related to retention since there were insufficient studies in our sample reporting follow-up sessions after initial training. In order to understand the long-lasting effects of neurofeedback future studies should incorporate follow-up sessions in their study designs.