Neurofeedback (NF) is a training method by which real-time feedback of brain activity, typically an EEG parameter, is delivered to the subject to promote voluntary control of brain activity. The subject has electrodes attached to the head, and the measured EEG parameter is converted to a sound or visual stimulus, which is then fed back to the subject. The main NF protocols for patients with attention deficit hyperactivity disorder (ADHD) are the training of frequency bands and the training of slow cortical potentials (SCPs). Frequency band NF targets tonic aspects of activation by promoting learning to reduce or to enhance activity of defined frequency bands. SCP training targets the phasic regulation of cortical excitability by learning to generate negative and positive shifts of cortical activity. SCPs originate in the apical dendritic layers of the neocortex and reflect synchronized depolarization of large groups of neuronal assemblies. According to Birbaumer’s threshold regulation model of cortical excitation [
Although frequency band training is the most common form of NF for ADHD, recent research no longer supports the presumption that increases in theta power, reductions in beta power, or the theta/beta ratio is a reliable ADHD marker and, in consequence, compelling targets for NF [
In recent meta-analyses of NF efficacy for ADHD [
The transfer of learning in NF with respect to everyday life situations is hypothesized to be better practiced in the transfer condition than in the feedback condition. In the transfer condition, the subject has to modulate the NF parameter without the aid of a feedback stimulus. The transfer condition is hypothesized to be closer to everyday life situations as compared to the feedback condition, where continuous performance feedback is available [
Neurofeedback for ADHD has mainly been perceived as an alternative for stimulant medication, but the combined effects of medication on NF learning are unknown. In several ADHD NF studies, MPH has been permitted in constant dose [
Although a great deal of evidence suggests that EEG activity is associated with age [
One challenge in analyzing training studies across multiple sessions is that the training performance variability varies considerably not only across time within a single subject but also across multiple subjects, which compromises conventional basic statistical methods, where correlations between observations are often obstructive. For that reason, we opted for a mixed-effects modeling approach. One major advantage of mixed-effects modeling is that it does not assume independence among observations and is to some degree more robust with unbalanced data than basic multivariate analysis (36).
In this study, we analyze NF learning in children and adolescents with ADHD. The major research question of this paper is (1) whether, and to what degree, both subject-specific (e.g., age or IQ) and treatment-related factors (e.g., school versus clinical treatment setting) may be related to NF learning within and across sessions, (2) whether NF learning differs in feedback and transfer conditions, and (3) whether within-session analysis can contribute additional information to cross-session analysis.
Subjects were recruited in outpatient clinics, by referral of clinicians, in parent self-aid groups, and at schools. Forty-four subjects, of whom 33 had a clinical ADHD diagnosis before entering the study, were included. See Table
Description of participants.
Total | With MPH | No MPH | |
---|---|---|---|
48 | 16 | 32 | |
Male/female ( |
27/21 | 12/4 | 15/17 |
Clinical setting ( |
26 | 10 | 16 |
School setting ( |
22 | 6 | 16 |
Intersession interval (days) | |||
Clinic | 4.1 ± 1.9 | 4.1 ± 8.3 | 4.4 ± 6.3 |
School | 4.8 ± 1.1 | 4.8 ± 3.2 | 5.3 ± 7.7 |
Age (years) all | 11.2 ± 2.2 | 10.9 ± 2.4 | 11.4 ± 2.0 |
MPH dosage (mg) | 24.5 ± 15.1 | 23.6 ± 15.0 | 0 |
MPH intake duration (years) | 2.3 ± 2.5 | 2.4 ± 2.5 | 0 |
Estimated IQ | 109.5 ± 14.8 | 109.9 ± 14.7 | 109.1 ± 15.3 |
Clinical ratings before training | |||
DSM-IV C-3 P (T-scores) | |||
Inattention | 67.8 ± 5.8 | 63.3 ± 6.4 | 67.5 ± 8.2 |
Hyperactivity/impulsivity | 64.9 ± 8.4 | 66.1 ± 6.2 | 59.6 ± 9.1 |
DSM-IV C-3 T (T-scores) | |||
Inattention | 65.7 ± 6.1 | 61.7 ± 7.0 | 64.2 ± 5.3 |
Hyperactivity/impulsivity | 63.0 ± 7.9 | 59.96.9 | 62.9 ± 8.3 |
Clinical ADHD diagnosis (yes/no) | 33 | 16/0 | 17/15 |
C-3 P/T: Conners 3 parent/teacher ratings (DSM-IV indices); MPH: methylphenidate.
Inclusion in the study required written consent by both the child and parents. The study was approved by the local ethics committee. Age ranged from 8.5 to 16.5 years. Inclusion in the study was based on clinically relevant scores in the German version of the Conners 3 parent and Conners 3 teacher rating scales [
Medication with methylphenidate (MPH) was allowed if the dose was kept stable over the full treatment time, including three months before the first assessment. For children taking MPH, teacher and parent ratings had to be based on the behaviour on medication. Exclusion criteria were estimated IQ ≤ 80 (short form of the German WISC-IV [
Parents and teachers rated the child’s behaviour on the Conners 3 scales and the Behaviour Rating Inventory of Executive Function (BRIEF) [
Study design.
NF was provided using a commercially available mobile training device (THERA PRAX; neuroConn GmbH). Double sessions consisted of four blocks, each containing 40 trials (see Figure
Setup of the SCP-NF. Feedback/transfer condition: condition where a feedback stimulus is (feedback) or is not (transfer) visible. Deactivation task: generation of positive potential shifts. Activation task: generation of negative potential shifts. 1 double session consists of 4 blocks with 40 trials each, each block including feedback and transfer conditions and deactivation/activation tasks as illustrated (pictures by Ilmenau, neuroConn GmbH).
The participants’ EEGs were recorded at electrode Cz, referenced to the right mastoid electrode (ground was left mastoid) shunted over a 10 kOhm resistance (impedance < 20 kOhm; sampling rate was 512 Hz). The EEG amplifier (THERA PRAX, neuroConn©) used a low-pass filter of 40 Hz. Filtering of the SCPs was performed from 0.01–40 Hz with a two-way least-squares FIR filter. Preprocessing was performed with MATLAB and EEGLAB. Processing of the SCPs (DC—2 Hz) was performed from channel Cz-A2 for each sample point and displayed on the trainer screen. The maximal time delay until the patient saw the feedback of the NF parameter was about 110 ms. Display of the change in mean amplitude with respect to the pretrial baseline was fed back by the vertical movement of the feedback stimulus, whereas its horizontal position corresponded to the time axis. Trials were baseline corrected (the mean amplitude of the pretrial baseline was subtracted from each data point of the SCP amplitude) and then averaged. Since we frequently observed muscle activity in the first second of the trial, we only incorporated the last 6 seconds of the recording in the active trial. As regression-based artefact correction procedures did not yield reliable results, we applied a strict artefact removal procedure, where after manual artefact rejection, baseline-corrected trials were rejected if their amplitudes exceeded ±100 mV or their gradients exceeded 50 mV between two data points.
Four separate models were analysed to predict performance in the feedback condition and transfer condition either across or within sessions. Statistical analysis was performed with a linear mixed-effects (LME) regression [
Effects considered for statistical analysis.
Model specifications | Measure |
---|---|
Time | |
Cross-session model | Double session number (15 double sessions) |
Within-session model | Bin number: 10 bins per session. The mean amplitude of baseline-corrected trials was averaged across sessions and then averaged across the ten equally spaced units (bins). |
Condition type | |
Feedback (FB) | Continuous performance feedback stimulus visible |
Transfer (TR) | Performance feedback, delayed |
Tasks | Deactivation (generation of positive potential shifts of SCPs) versus activation (generation of negative potential shifts of SCPs) |
Intersession interval | Days passed between training sessions |
Age | In years (continuous variable in the model, only for visualization in plots dichotomized into younger and older age classes) |
MPH | Being on constant stimulant medication (methylphenidate), factorized into yes versus no |
Stimulants intake duration | Years of MPH intake |
Dosage of stimulant medication | Methylphenidate (MPH) in mg |
Sex | Factorized into female versus male |
IQ | Estimated IQ (WISC-IV short form) |
Setting | Factorized into school setting versus clinical setting |
Severity of ADHD symptoms | |
Preexisting ADHD diagnosis | Clinical ADHD diagnosis before entering the study factorized into yes versus no |
Artifact rate | Percentage of rejected trials within a session |
The statistics of the best model fit for each of the four models to predict NF performance are presented in the following sections. We will call performance progress in each condition “feedback learning” and “transfer learning,” respectively.
As shown in Table
Results for NF learning with respect to condition (feedback/transfer) and time (cross-/within-session).
Cross-session learning | Within-session learning | |||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
Feedback | Transfer | Feedback | Transfer | |||||||||
B | CI | B | CI | B | CI | B | CI | |||||
Intercept | −3.95 | −5.58 to −2.44 | −1.06 | −2.72–0.68 | 0.218 | −2.65 | −3.86 to −1.40 | 1.27 | −0.38–2. 68 | 0.095 | ||
Session | −0.00 | −0.17–0.16 | 0.994 | 0.03 | −0.18–0.21 | 0.797 | ||||||
Bins | −0.29 | −0.40 to −0.18 | −0.40 | −0.61 to −0.20 | ||||||||
Task | 1.66 | −0.04–3.31 | 0.061 | −2.40 | −4.50 to −0.56 | 2.79 | 2.14–3.41 | −2.92 | −4.34 to −1.27 | |||
Age | −0.49 | −1.28–0.27 | 0.214 | −0.59 | −0.94 to −0.24 | −0.62 | −1.20 to −0.16 | −0.40 | −0.78 to −0.00 | |||
MPH | 1.18 | −1.44–3.66 | 0.384 | 4.11 | 1.34–7.13 | 1.17 | −0.57–2.87 | 0.176 | 1.71 | −0.03–3.40 | 0.051 | |
IQ | −0.08 | −0.14 to −0.02 | −0.07 | −0.13 to −0.01 | ||||||||
Session: task | 0.15 | −0.02–0.34 | 0.123 | 0.20 | −0.02–0.4 | 0.072 | ||||||
Bins: task | 0.33 | 0.07–0.56 | ||||||||||
Session: age | −0.01 | −0.09–0.08 | 0.858 | |||||||||
Task: age | −0.42 | −1.33–0.43 | 0.342 | −0.40 | −0.71 to −0.08 | |||||||
Session: MPH | −0.05 | −0.32–0.21 | 0.745 | −0.27 | −0.63–0.06 | 0.124 | ||||||
Task: MPH | −1.56 | −4.53–1.91 | 0.308 | −4.35 | −7.54 to −1.13 | −0.79 | −1.98–0.39 | 0.177 | ||||
Age: MPH | 0.38 | −4.53–1.91 | 0.538 | −0.73 | −1.45–0.09 | 0.066 | ||||||
Session: task: age | −0.01 | −0.12–0.09 | 0.804 | |||||||||
Session: task: MPH | 0.12 | −0.20–0.46 | 0.464 | 0.39 | 0.05–0.76 | |||||||
Session: age: MPH | −0.15 | −0.28 to −0.01 | ||||||||||
Task: age: MPH | −1.41 | −2.70–0.019 | 0.86 | 0.36–1.35 | ||||||||
Session: task: age: MPH | 0.32 | 0.16–0.47 | ||||||||||
41.776 | 49.174 | 17.870 | 32.477 | |||||||||
7.214 | 8.501 | 9.093 | 13.320 | |||||||||
0.0734 | 0.1325 | |||||||||||
0.05569 | 0.1231 | |||||||||||
−0.455 | −0.620 | −0.585 | −0.735 | |||||||||
Observations | 1400 | 1380 | 959 | 959 |
Mixed-effects model results for NF learning. The dependent variable is mean amplitude (
Visualization of effects moderating cross-session NF learning in the feedback condition. The dependent variable is mean amplitude (
As shown in Figure
The final model for within-session learning for the feedback condition included subject as random intercept (
Visualization of effects moderating within-session NF learning in the feedback condition. The dependent variable is mean amplitude (
A three-way interaction between task, MPH, and age resulted in the best model fit (
The final model for cross-session learning for the transfer condition included subject as random intercept (
Visualization of effects moderating cross-session NF learning in the transfer condition. (a) Interaction effect between session, task, and MPH. Transfer condition: no continuous feedback stimulus visible. Task. Deactivation: generation of positive potential shifts. Activation: generation of negative potential shifts. MPH: being on constant methylphenidate medication. (b) Age effect plot.
The final model for within-session learning in the transfer condition included subject as random intercept (
Visualization of effects moderating within-session NF learning in the transfer condition. (a) Interaction effect between bin number and task. Transfer condition: no continuous feedback stimulus visible. Task. Deactivation: generation of positive potential shifts. Activation: generation of negative potential shifts. (b) MPH effect plot. MPH: being on constant methylphenidate medication.
Thus, NF learning in the transfer condition took place in the activation task rather than in the deactivation task. Moreover, being on constant methylphenidate medication was associated with a more positive mean amplitude (see Figure
We also analyzed whether NF learning was associated with the number of trials rejected due to artifacts by performing separate models for within and cross-session learning that included artifact rejection in the models. The mean artifact rate was 29.1% (±17%). The inclusion of the artifact rate did not yield a significantly better model fit for either condition.
To explore the number of subjects showing the desired learning slope in cross-session NF learning, models for both the feedback and transfer conditions were calculated separately and the subjects’ random slopes were extracted to determine the individual learning performance for each task. Successful NF learning was defined by a negative slope in the activation task or a positive slope in the deactivation task. Subjects presenting both a positive slope in the deactivation task and a negative slope in the activation task were labelled “successful regulators.” In the feedback condition, 20 learners (41.7%) in the activation task, 23 learners (47.9%) in the deactivation task, and 10 subjects (20.8%) were classified as successful regulators. In the transfer condition, 23 subjects (47.9%) were classified as learners in the activation task, 23 as learners in the deactivation task (47.9%), and eight as successful regulators (16.7%).
This paper addresses the lack of NF studies in ADHD that map learning in NF and control for both treatment-related effects, such as setting and time frequency, and subject-related effects, such as IQ and stimulants. It presents the groundwork for measuring treatment specificity [
Children on constant MPH showed stronger performance increments across sessions with increasing age (age range between 8.5 and 16.5 years). In contrast, children who did not take MPH showed less pronounced potential shifts than when on constant stimulant medication. For these children, learning was negatively moderated by age, albeit the generation of potential shifts was still in the desired relative direction (mean amplitude in the activation task was more negative than that in the deactivation task).
Similarly to the cross-session NF model, performance was also interacting with age and MPH of comparable direction and strength. In contrast to cross-session analyses, children generated negative potential shifts within sessions irrespective of task and time. However, the generation of potential shifts remained in the desired direction (mean amplitude in the activation task was more negative than that in the deactivation task). Thus, children produced progressively more negative potential shifts throughout a session, irrespective of whether the task demanded positive or negative potential shifts. Since moderators of learning have been rarely examined in SCP-NF before, these findings are difficult to explain in the context of previous research. It is open to speculation whether this finding might reflect the time required to fully mobilize attentional resources within a session. The added value of within-session analyses in the feedback condition relies here on the possibility that two consecutive training sessions of NF might not necessarily be too tiring for children and adolescents with ADHD; on the contrary, our findings might even indicate that subjects need time to immerse themselves in the training scenario if they are to tap into the full potential of the training, especially with respect to the activation task. Thus, it might even be recommended to perform trainings in the form of double sessions.
The NF literature offers little help in interpreting these opposite findings with respect to medication and age (feedback learning across sessions was positively associated with age for children with stimulants, but negatively associated with age for medication-free children). Previous NF studies allowing MPH have not included these factors as covariates for learning together [
Taken these results together, it appears that feedback learning may become easier and faster with MPH and increasing age. Therefore, it might be more beneficial for older children taking stimulants to increase the proportions of transfer trials earlier in training sessions than for younger children not taking stimulants. Older subjects taking stimulants might benefit earlier from generalizing effects of the acquired NF skills. In contrast, younger children without MPH might need more training sessions and more feedback trials to consolidate the NF skills.
Children with a higher estimated IQ generated more negative potentials, irrespective of other effects such as time, task, age, and stimulants. This finding was expected and is supported by another study showing that the CNV, one form of a SCP reflecting cognitive mobilization, was positively associated with IQ [
Transfer learning was especially challenging, as shown by potential shifts that were smaller than those in feedback learning. As no continuous performance feedback is available during the transfer condition, regulating attention becomes more difficult. Furthermore, and in line with Strehl et al. [
Within sessions, transfer learning took place only in the activation task but remained unchanged in the deactivation task. Thus, subjects managed to improve the voluntary upregulation of attention within a session, while the voluntary downregulation of attention remained stable. It is difficult to interpret this finding. As with the within-session feedback learning, it might have taken the subjects some time to fully mobilize attentional resources within a session. Currently, no study on SCP-NF has reported results on within-session learning (but see 22, 24, 48, 49 for within-session analyses for frequency band NF in ADHD). Thus, further research is needed to map learning within sessions and to fully understand its interdependency with learning across sessions.
In transfer learning both within and across sessions, age was negatively associated with the mean amplitude irrespective of time or session number. This association was probably related to larger proportions of fast frequencies as a function of age [
Taken these findings together with respect to condition, task type, time window and subject-related or treatment-related factors, age and stimulants were the dominant moderators of learning: in medicated children, age was positively associated with NF performance while being negatively associated in nonmedicated children—for both within- and cross-session analyses. In contrast, transfer performance across time was only moderated by MPH and only when considered learning across sessions, but not within session. In this study, transfer and feedback trials were mixed within one block and the number of transfer trials increased across sessions.
Neither dosage nor duration of stimulant intake predicted learning. However, we cannot exclude any general effects of dosage and intake duration on learning, since dosage and duration of stimulant intake did not vary by amounts that might have led us to expect possible moderating effects. Clinical symptoms or severity rated by parents and teachers did not moderate learning. This was unexpected, since we had hypothesized a more severe initial impairment of attention to be reflected in weaker overall NF performance. However, clinical severity might have not been linearly associated with performance but might have been moderated by a threshold of relevant impairment; we did not investigate this issue. The artifact reduction rate has been shown in previous studies to improve over time, possibly as a nonspecific effect of the treatment helping children learn to sit still [
It was not surprising that setting was not associated with NF learning, as NF learning should not be affected by the training environment; however, differential setting effects on NF learning have never been tested directly before, so our study is the first to provide empirical confirmation of this common assumption. Likewise, intersession interval has rarely been examined as an effect on NF learning or clinical improvement [
By employing a mixed-effects modeling approach, we expected to achieve a more realistic mapping of NF learning in ADHD than other statistical models, such as multivariate analysis of variance (MANOVAs). First, results achieved by MANOVAs are very sensitive to outliers, and furthermore, results can easily be biased by unbalanced datasets and missing data. Mixed-effects models can deal with these impediments to a certain extent. A major advantage of our statistical approach when drawing conclusions about the usefulness of MPH for NF learning is that independence amongst observations is not a necessary precondition; performance variability can be accounted for both within a subject across sessions and between subjects. One limitation of this approach may be the lack of current consensus whether and if so by what degree it is possible to rely on
The study did not include follow-up or booster sessions. Although there is evidence that SCP-NF performance can be maintained at least up to two years [
Given the complex interactions in our results which have not been shown before, we conclude that mixed-effect modeling is an appropriate approach to analyze NF learning. We therefore suggest this approach for future research to reach a better understanding of the mechanism of NF learning and treatment specificity.
Daniel Brandeis reports serving as an unpaid scientific advisor for an EU-funded SME study on neurofeedback. The other authors have no conflicts to declare. This funding did not lead to any conflict of interests.
The authors thank Dr. Alex Roth for his support with linear mixed modeling statistics. This project has been funded by the Swiss National Science Foundation.
Figure S1: visualization of cross-session NF learning in the feedback (A) and transfer conditions (B). The dependent variable is the difference between mean amplitude (visualization of cross-session NF learning in the feedback (A) and transfer conditions (B)). For raw data, see scatter plot under each effects panel, fitted with a fixed linear regression based on the same factors as in A. A: interaction plot for the fixed effects session number, MPH, and age. B: interaction plot for the fixed effects session number and MPH. Session number: 15 sessions in total. Condition: deactivation: generation of positive potential shifts. Activation: generation of negative potential shifts. MPH: being on regular methylphenidate medication (yes versus no). Condition: feedback: feedback stimulus visible. Transfer: no feedback stimulus visible. A and B: for visualization age is subdivided into two age classes (8–12 and 13–16 years), but preserved as a continuous variable in the original model. Table S2: results for linear mixed effects models for cross-session NF learning with a condition type (feedback/transfer) in titles. Table S3: ANOVA results of models predicting NF performance in the feedback condition. Table S4: ANOVA results of models predicting NF performance in the transfer condition.