Review on Frequency Neurofeedback on Autism Spectrum Disorder: Overview, Efficacy and Research Direction

The current author searched for scientific articles in Google Scholar and specific databases such as PubMed, Ovid MEDLINE, EMBASE, ERIC, and CINAHL using the key words: neurofeedback, neurofeedback training, neurotherapy, NF, NFT, NFB, autism, ASD, autism spectrum disorder. Search queries were adopted using Boolean operators AND and OR to locate studies. Additional references were identified in key literature reviews e.g., Coben et al. (2010), Marzbani et al. (2016), Holtmann et al. (2011), van Hoogdalem et al. (2020). English was chosen as the search language. Experimental articles up to May 2025 were analyzed.

Studies were included in this review if they were: a) peer-reviewed; b) written in English; c) had a reasonable sample size; and d) were a randomized and/or matched controlled trial. There were some studies that were excluded, including: Zivoder et al. (2015), where the sample was small (n=10) and the authors failed to report data even though findings were discussed; Darling (2007), where the paper was not peer-reviewed and his paper had a small sample size (n=6); and Mohammadi et al. (2019), which was not in English.

Tables 1 shows ten studies reported across ten published articles with one article reporting two studies (Pineda et al., 2008) and one research group reporting follow-up data in a separate article (Kouijzer et al., 2009a, 2009b). The first published trial was in 2002 by Jarusiewicz. There is less than one study every two years being published in this field, which begs the question, why is the publication amount so low? FNF is still a relatively obscure and esoteric clinical therapy and is only beginning to attract research focus. A major barrier is the cost of equipment and training; FNF is a relatively complex therapy and to conduct research requires an experienced clinical researcher to govern the protocols using equipment and licences that cost at least $5,000USD. It entails a considerable commitment from the subject e.g. 10 hours plus of training time. Moreover, there is a growing number of studies on alternative neurofeedback methodologies such as Slow Cortical Potentials (SCP), fMRI neurofeedback, and infra-low neurofeedback (ILF), which take focus from FNF as an intervention worthy of study.

A glaring shortcoming of these studies is consistently small sample sizes, that equates with low-powered studies. Button et al. (2013) argue that low powered studies impact in three ways, including a) the low probability of finding true effects; b) the low positive predictive value when an effect is claimed; and c) an exaggerated estimate of the magnitude of effect. Points a) and b) reduce the probability of a true positive finding, and c) suggests that the effect is misleading because the finding is an outlier or the ‘winner’s curse’. The ‘winner’s curse’ suggests that the first promising finding (e.g. Jarusiewicz, 2002) is an exaggerated positive result, and any attempts to replicate are difficult, presumably as they are not outliers. However, Jarusiewicz’s study has been replicated several times, and Coben and Padolsky (2007) demonstrated a stronger effect with half the amount of FNF training using QEEG-derived protocols. This does not support the notion of a ‘winner’s curse’ in the FNF research.

Larson and Carbine (2017) recommend calculating sample sizes before engaging in sampling, or at least calculating correlations between pre- and post-test measures to accurately calculate future sample sizes. This ensures that studies are sufficiently powered, reduce the probability of a false negative finding (‘Type II’ error), and ensure that effect sizes are not inflated. This has not happened consistently in EEG or ERP studies, and this recommendation should be followed when studying FNF on autistic populations.

There is also a possibility of publication bias, since only a handful of studies have been published since 2002. The publication rate has been low, and this may be due to many null findings not being published as they are unremarkable to scientific journals. Begemann et al. (2016) conducted a meta-analysis of five studies relating to FNF on autism and “showed a large superior effect of 0.85 (p=.003, 95% CI=0.29 to 1.40). Heterogeneity was moderate …, publication bias was not indicated.” They added: “Four studies combined showed a large superior effect for neurofeedback compared to waiting list or skin conductance therapy (ES 0.80, p=.029, 95% CI=0.08 to 1.52) … Pineda et al (2008) found a large superior effect of neurofeedback over placebo treatment (ES 0.96, p=.039)” (p. 25). Moreover, Begemann et al. (2016) report a fail-safe N_R for ASD general symptomatology of 18 and a fail-safe N_R for passive treatment of 11. This suggests that many null finding studies (11 to 18 times greater than positive finding studies) are required to negate the positive effects reported in any given published study.

Based on the Begemann et al. (2016) review, assuming an effect size of .80, significance of .05 and power of .8, the estimated sample size per group is 26 (N=52). Researchers should account for dropout rates across groups, and aim for a final, eligible sample size of 26 and not an initial larger sample size at the beginning of a study. A participant of a FNF study is required to engage in several sessions each week, over a few months, which is a large commitment, and a moderate dropout is expected. The dropout group should also be compared with completing subjects to assess varying characteristics of these groups, that may influence the dependent variable. Carrick et al. (2018) failed to analyze completing participants compared with non-completing participants, which is a significant analytical flaw of this study (see limitations in Table 1).

A trend in the research is that autistic groups reported in Table 1 are usually children (4-17 years) who are verbal with an IQ score >70. Males have been selected in greater proportion in these studies, usually on a ratio of 3:1 or 4:1. Autism is diagnosed more often in males, which explains the higher ratio of males to females studied. The incidence rates of autism in males and females are often attributed to genetics, however there is some argument of the under diagnosis of females, due to a varying presentation, and diagnostic bias towards males (Tsirgiotis et al., 2023). In spite of these arguments, future research could focus on factors such as biological sex, IQ level (including intellectual deficit, ID, versus non-ID) and verbal abilities (i.e. non-verbal, minimally verbal, verbal).

To summarise, the findings suggest a generalization to young males under 18 years, who are verbal with an IQ above 70. This would capture a reasonable proportion of the population with autism but does not generalize to adults with autism, or autistic females of any age.

A major area if evaluating the efficacy of FNF is the type of methodological designs employed in the research. The factors to consider include selection process and allocation (i.e. self-selected, randomization), blinding (i.e. participant and parent only, or double blinding, including FNF technician) and type of control (i.e. waitlist, active, matched). Controlling nonspecific factors is very important, such as: contextual (e.g. subject is engaging in an experiment and expects to improve, positive regard for technician, EEG display is perceived as scientific and therefore efficacious); attention effort (e.g. participant watches a screen for 30 minutes, 30-50 times and improves due to focussing for long periods of time); repeated reinforcement (e.g. technician gives verbal reinforcement, audio-visual feedback unrelated to EEG data); and, specifically with autism, the routine of attending regular sessions, which can satisfy the ‘need for sameness’ or a routine. The gold standard in biofeedback is a double-blind randomized sham or active controlled design (van Doren et al., 2019). The active control group receives cognitive training or EMG biofeedback to control for nonspecific effects e.g. placebo.

Four studies used the gold standard of a blinded randomized sham control (Pineda et al., 2008, studies 1 and 2; Carrick et al., 2018; Wang et al., 2024) with another study controlling for nonspecific effects by adopting a typically developed control group (Datko et al., 2018). The Carrick et al. (2018) study is problematic because the dropout rate was greater than 50% in the experimental and control groups with no reported data analysis of dropouts versus completing participants across the groups. If the experimental group dropouts failed to respond to FNF, then the data from this subgroup may have negated the effect reported in the study. Datko et al. (2018) is also problematic because the researchers could have incorporated sham controls for both ASD and typically developed groups. The two Pineda et al. (2008) and the Wang et al. (2024) studies are consistent in findings, that is, a reduction in the ATEC Sensory/Cognitive Awareness subscale.

It is worth mentioning that Wang et al. (2024) used a technician delivered and home-based AI-driven device that delivered audio-only feedback, with a reduction in core autistic traits as measured by the ATEC.

One of the strongest features of this body of research is the use of standardized psychometric measures. Researchers mostly employed the Autism Treatment Evaluation Checklist (ATEC), which is available free of charge and has been validated as a treatment outcome measure for autistic populations (Rimland & Edelson, 1999). The ATEC has an advantage of providing normative scores based on age, and accounts for change based on maturation. If a study uses the ATEC and has a one-year follow-up methodology, then actual change scores must be compared with change scores due to maturation alone.

All studies have used a pre- and post-therapy methodology, spanning over a few months to six months, except one study that reported post-training follow-up data (Kouijzer et al., 2009a, 2009b). Kouijzer (2009b) reported that any immediate behavioral and cognitive improvements from FNF were maintained (and some participants improved even more) at a 12-month follow-up. These findings are consistent with van Doren et al.’s (2019) meta-analysis which reported that ADHD symptoms reduce even further from post-FNF treatment to 12-month follow-up, due to the mediating factor of sleep improving.

Another strength of this research is the use of specific, standardized psychometric and diagnostic measures of social functioning (e.g. SDS), IQ (e.g. WISC), attention (e.g. TOVA), autism diagnostics and traits (e.g. ADI, CARS), emotion regulation (e.g. ERC), anxiety (e.g. Spence Anxiety), medical diagnostics (e.g. fMRI, QEEG) and cognitive functioning (e.g. BRIEF). Whilst this is a relatively comprehensive use of standardized psychometrics, there are some areas that are yet to be measured, that are important for people with disability, such as: quality of life, education and learning, employment, activities of daily living (ADLs) and mood (i.e. depression). Researchers should aim to use measures that are sensitive to change and applicable to neurodivergent populations (e.g. ASQoL, ASC-ASD).

A further strength of this field is the use of multiple sources of data to correlate the effects of FNF and pinpoint a mechanism. For example, Datko et al. (2018) correlated psychometric measures (e.g. autistic traits, ATEC, social functioning, SDS) with diagnostic data (e.g. ADOS, ADI) and fMRI. Future studies should aim to target an area of dysfunction (e.g. social deficits), hypothesise an area of training that would reduce a dysfunction (e.g. mu rhythm using C4-A1), and measure brain functioning (e.g. fMRI) or a biomarker (e.g. event-related potential, ERP) that reflect those changes. The ultimate test of FNF efficacy is: a) EEG changing in the direction predicted by the protocol (i.e. increased mu suppression, increased SMR); b) emotional, cognitive and/or behavior change that is correlated with EEG change; and c) maintenance of these changes following the withdrawal of the treatment i.e. self-regulation in the absence of brain-derived feedback.

A potential weakness of these studies is the use of parent-only ratings which are defined as ‘most proximal’ and ‘least blinded’ ratings. Cortese et al. (2016) showed that the experimental effects of FNF for ADHD diminished to non-significance when teacher ratings were analyzed across multiple studies, since teacher ratings are presumably ‘less proximal’ and ‘probably blinded’. This phenomenon was also replicated with sham controls with blinded raters. Van Doren et al. (2019) argue against the ‘proximal-blinded’ concept suggesting that parents rate different cognitive and behavioral elements of ADHD than a teacher, or ADHD behaviors are rated in a different context e.g. home, community. Furthermore, Table 1 shows that Pineda et al. (2008) and Carrick et al. (2018) used sham controls and still found an effect, whilst Kouijzer et al. (2013) used a double-blinded procedure for FNF and skin conductance groups (both groups were identically prepared and were unaware of the true feedback they were receiving). An effect was demonstrated with blinded parents, however the separation of experimental participants into regulators versus non-regulators is slightly dubious. Seven out of 13 subjects in the FNF group were identified as regulators, suggesting that approximately 50% of cases will respond to FNF when using standardized protocols. Outcomes were maintained at a six-month follow up.

There is a call for standardization of training in FNF research both in terms of protocols and delivery of training (Arns et al., 2020). In regards to delivery of training, Table 1 shows a wide variety of approaches to training, including auto-thresholding (Pineda et al., 2008) and/or high reward delivery (Goodman et al., 2018). The problem with auto-thresholding or high reward delivery is that a constant reward is delivered regardless of EEG output, which violates the ‘successive approximation’ approach required in FNF to train EEG frequencies to normative or desirable power limits.

There is also a significant issue associated with standardized protocols. Coben and Padolsky (2007) showed that using QEEG-derived protocols increased the magnitude of the effect with half the number of sessions when compared with Jarusewicz’s original study. That is, if standardized protocols are utilized, then the effect will be smaller than if QEEG-derived protocols are adopted. There is the additional dilemma that autism is associated with a spectrum of behaviors, which may explain the wide variety of protocols across studies, and the advantages of individualized protocols based on QEEG data. It is disadvantageous to standardize protocols and expect an effect size that is truly reflective of the efficacy of FNF. In the ADHD field there is a relatively standardized approach to target fronto-central slowing (Fz-A1, Cz-A1) or hyperkinetic behavior (SMR training), and even in that field there is a movement towards ‘precision medicine’ (Arns et al., 2014).

Researchers have also focused on equipment and their characteristics. In my opinion, the factors that matter in FNF are not the equipment and their idiosyncrasies. What matters is the montage, how much and in what way reward is delivered, duration of training, number of training sessions to achieve a treatment response, training density (e.g. frequency of sessions per week), the immediacy of the feedback, type of feedback (i.e. audio, tactile, visual), and the bandwidths rewarded and inhibited. These factors are considered in a recent paper by Bazanova et al. (2025). This is a matter of process, which is the essence of FNF. The technician regulates settings in a reflexive manner to guide the brain to function efficiently and optimally during training. In operant conditioning terms, ‘successively approximate’ or ‘shape’ to efficient electrical activity in specific frequencies. This requires the technician to have a minimum level of experience and competency. There are competency practice standards for FNF training (Biofeedback Certification International Alliance, BCIA, ‘Blueprint of Knowledge Statements’, 2018), however it is unlikely these standards were met in the extant research. A more detailed analysis and discussion of FNF research on healthy subjects is provided elsewhere (Rogala et al., 2016).

Finally, the most common protocol was to reward alpha frequencies (i.e. 8-13 Hz) with an aim of targeting the mirror neuron system (MNS) and training mu suppression to impact brain functioning when observing social interactions or movement. This connection was elegantly demonstrated by Pineda and his colleagues (see Courellis et al., 2019), in which they down trained mu rhythm during emotion-focused FNF sessions, with increased mu suppression in key networks with the brain approaching a typically developed child. I will discuss the importance of Pineda et al.’s (2008) research in the Theory section. The point here is that researchers could expand on our understanding of causal mechanisms by training frequencies other than alpha and testing hypotheses other than Theory of Mind (ToM).

All studies show a positive effect when FNF is compared with controls, under random or matched assignment, for pre- and post-testing. A first impression suggests that even with some publication bias this represents a positive conclusion of the evidence base. Moreover, publication bias against null-hypothesis findings may suggest more about a tendency to support accepted psychological theories (and reject findings that weaken our belief in these) than a mere aversion to null findings (Ferguson & Heene, 2012). FNF straddles psychological and neuroscientific theories (in other words, biological psychology), and the research has pursued simple and pragmatic outcomes such as reducing social deficits and behaviors of concern, instead of testing psychological theories like a ToM.

If we assume that publication bias is a minor problem in this field, then we can consider the efficacy level of FNF on autism. Coben and Padolsky (2007) write: “Our study may be the first step in establishing a Level 3 criteria rating of neurofeedback as probably efficacious in the treatment of ASD. We replicated another controlled study (Jarusiewicz, 2002). A broader range of outcome measures confirmed the reduction of ASD symptomatology following neurofeedback” (p.18). There have been further studies published since Coben and Padolsky (2007), which means the overall efficacy level of FNF is re-considered here.

To begin with, FNF autism research cannot be assessed using the stricter guidelines outlined in Arns et al. (2020) because there are no systematic reviews completed in the last two years. The last systematic review was completed by Begemann et al. (2016). Therefore, we can only assess this body of research using the lower standards outlined by La Vaque et al., which outlines an efficacy framework for psychophysiological interventions with five levels of efficacy. The first level is anecdotal or case studies through to the fifth level of “Efficacious and Specific. The investigational treatment has been shown to be statistically superior to credible sham therapy, pill, or alternative bona fide treatment in at least two independent research settings” (p.280).

All studies in Table 1 show positive findings with some inconsistent results. There are two studies that meet the gold standard methodology, including Pineda et al. (2008) and Wang et al. (2024). Both studies employed a double blinded randomized sham-controlled methodology. Both studies targeted mu rhythm using a suppression approach, with results consistent across the studies i.e. reduced scores on the ATEC Sensory/Cognitive Awareness subscale. The Wang et al. showed a reduction in the full scale ATEC score that approached significance (p=.082). The Wang et al. (2024) study failed to employ the video conditions (e.g. Hand, Crayon, Social), utilized by Pineda et al. (2008), who showed that mu suppression was higher in the treatment group when watching these videos following FNF compared with controls. Level 5 ‘efficacious and specific’ cannot be claimed, rather Level 4, ‘efficacious’ only, because this component was not replicated by Wang et al. (2024).

A weak aspect of the research has been the practice of describing autism according to the DSM-5 nomenclature and then demonstrating a reduction in autistic traits without any connection to a theory that may explain the disorder, or demonstrating neuro-mechanisms that underlie change in EEG and other factors e.g. behaviors of concern. The concept of understanding the mechanisms of FNF has been recommended in the ADHD research (Arns., et al, 2014). The clear leader in training specific protocols that link theory and neurophysiological mechanism is Pineda and his colleagues. They have shown that training mu suppression (8-13 Hz) during emotion-focused FNF training at C4, causes connectivity changes in the brain in key networks like the DMN and ToMN. This elegantly links to the psychological theory of ToM developed by Baron-Cohen et al. (1985). There have also been papers targeting executive dysfunction (first proposed by Minshew & Goldstein, 1998; Pennington & Ozonoff, 1996). Kouijzer et al. (2009a) showed that inhibiting theta and rewarding beta targeted under-connectivity in the anterior cingulate gyrus (ACC), recognised as a brain region that regulates cognitive and emotional processes associated with cognitive control and executive function (Bush et al., 2000). Kouijzer et al. (2009a) also write of the relationship between the activation of the default mode network (DMN)/ACC during task demand, which improves performance, and how ASD is associated with a hypoactivation of the DMN/ACC under attentional demand.

There has been no research on the third key psychological theory of autism called a ‘weak Central Coherence’ (Frith & Happe, 1994). Part of the problem may lie in quantifying ‘Central Coherence’ and in positing which brain networks are associated with deficits in this capacity.