We recruited 94 children diagnosed with ADHD (ADHD) and 23 healthy controls (HC). All participants were male and aged between six and nine. Mean and standard deviation of age were 7.0 ± 0.20 and 7.5 ± 0.28 in the ADHD and the healthy groups, respectively. All participants' parents gave their informed consent in accordance with the Declaration of Helsinki prior to the study. The Medical Ethics Committee of the Siberian branch of the Medical Science Academy approved this study. All patients fulfilled DSM-IV criteria for ADHD (American Psychiatric Association, 2013). No children with any comorbidity were included.

ADHD group was further categorized into the three qualifying types: (a) primarily inattentive type (n = 29), (b) primarily hyperactive-impulsive type (n = 30), and (c) combined type (n = 35). ADHD group overall differed significantly from HC in inattention (mean_ADHD = 3.72, mean_HC = 1.10, t = 12.45, p < 0.01) and hyperactivity/impulsivity (mean_ADHD = 2.42, mean_HC = 0.91, t = 11.97, p < 0.01).

First, groups were compared based on EEG, EMG, and psychometric characteristics (see section Psychometric Assessment). Then ADHD participants were randomly assigned to four groups for NFT to decrease TBR:

NFT consisted of 10 sessions, 16 min each. Despite the considerable amount of experience in employing NFT in ADHD, the optimal amount of time (and cost) of the treatment is still on debate. A typical course of NFT involves 30–40 sessions, which is very time consuming and expensive; particularly when considering that NFT is rarely covered by health insurance. Thus, there is a constant interest in developing a NFT protocol clinically beneficial in shorter time. Hillard et al. (2013) suggests that this may be achievable. The second reason for choosing fewer sessions was the inefficiency of standard TBR NFT, which led to that parents were reluctant to prolong NFT after 10 sessions. For a fair comparison, we did not used longer training for the novel protocols either, even if they were (more) efficient. Session length of 16 min was chosen due to our experience that 6–9 years old children with ADHD could not be engaged in a computer game for longer than ca. 16 min.

Participants were assessed pre-, post-, and 6-month after the NFT. Assessment consisted of psychometrics, behavioral rating scales completed by parents and teachers, as well as 19-channel EEG and forehead EMG recordings (Figure 1).

Psychometric assessment included Go/no-Go task (Avila et al., 2004) for assessing attention, delayed gratification task (Shoda et al., 1990) for assessing impulse control and Parent and Teacher Rating Scale interview (SNAP-VI) (Swanson et al., 2013) for assessing inattention, impulsivity and hyperactivity in behavior. The tasks have been performed as follows:

The clinical efficiency of NFT was measured as percent change in the reaction time and the number of missed stimuli in the Go-/no-Go task, as well as in the duration of delayed gratification at the post-NFT and the follow up sessions in comparison with the pre-NFT session.

We compared resting state EEG recordings of the two groups in 1 min eyes closed (EC) and 1 min eyes open (EO) condition at the pre-, post-NFT and the follow up sessions. Participants sat in a comfortable chair with a backrest and were instructed not to move their eyes during the recordings. EEG data was acquired using a Neuron-Spectrum-4 system (Neurosoft, Ivanovo, Russia) at a sampling rate of 1,000 Hz and with a bandpass filter between 0.1 and 50 Hz. The EEG electrodes were placed in accordance with the international 10–20 system: 18 electrodes were positioned in four regions: left fronto-central (LFC: Fp1, F3, C3, F7, T3,), right fronto-central (RFC: Fp2, C4, T4, F4, F8), left posterior-occipital (LPO: T5, P3, O1), and right posterior-occipital (RPO: T6, P3, O2), and signal was averaged within each area. Two reference electrodes were fixed at M1 and M2. The ground electrode was placed at Fp1 site. Impedance was kept below 5 kΩ. Data was anti-aliasing/low-pass filtered at 30 Hz.

The EEG data were analyzed using WinEEG software (Mitsar Co., St. Petersburg, Russia). Eye-blink artifacts were removed using Independent Component Analysis, then the recording was divided into 4.096-s-long epochs using a Hanning window (epochs were overlapped by 50%). Only records showing an amplitude above 10 μV² were included in the analysis. Analysis of EEG characteristics was carried out according to our previously published approach (Bazanova and Aftanas, 2010). Briefly, individual alpha peak frequency and alpha frequency band limits were determined at every electrode site for every participant using data from the last 30 s of EC and first 8 s of EO conditions. Frequency bands that were decreased in power by more than 20% during the EO vs. EC conditions were selected as individual alpha bands. Frequency showing the highest spectral amplitude within this band was identified as individual alpha peak frequency (IAPF). Lower and upper limits of the band [i.e., theta (TF) and beta (BF) transition frequencies] were determined where EO curve crossed the EC curve closest to IAPF. Theta and beta bands were defined between 3 Hz and TF, as well as between BF and 20 Hz, respectively. These adjusted band ranges were also used for the NFT. The alpha1 and alpha2 sub-bands were also adjusted: the alpha1 ranged from the lower limit to the individual alpha peak frequency, while the alpha2 ranged from the individual alpha peak frequency to the upper limit (Figure 2). The alpha amplitude suppression magnitude (ASM) in response to eyes open was measured as: ASM = log [(mean alpha power in EC condition – mean alpha power in EO condition) ^* 100%] / mean alpha power in EC condition). Alpha1 to alpha2 ratio (a1/a2) was also calculated.

EMG and EEG data were simultaneously recorded using Bosalb system, (NIIMBB SORAMN, Novosibirsk, Russia) with the sampling rate of 720 Hz. Feedback signal was recorded from Pz scalp electrode with a bandpass filter between 3 and 30 Hz. Monopolar reference electrode was located on the right ear, and the ground electrode was located at Fp1. Electrode impedance was kept below 5 kOhm. EMG was recorded using two Ag/AgCl surface bipolar electrodes with an effective diameter of 1.6 cm placed about 3–5 cm apart on the forehead. The EMG signal was amplified and filtered with 5–350 Hz bandpass filter.

Location of Pz was selected based on the following considerations: (i) alpha activity in the parieto-occipital area is generally proposed as a modulator of the attentional states (Ergenoglu et al., 2004); (ii) alpha amplitude in posterior areas are hypothesized to reflect top-down inhibition to suppress distracting information (Min and Herrmann, 2007); (iii) posterior alpha measures are exclusively used as predictors of individual differences (Bazanova, 2011; Tenke et al., 2013; Bazanova and Vernon, 2014); (iv) test-retest investigation had shown that alpha peak frequency, bandwidth, and amplitude are the least variable and the most reproducible at the posterior brain regions (Bazanova, 2011; Bazanova and Vernon, 2014); (v) due to the nature of alpha wave generation, the changes in alpha activity recorded in posterior scalp surface reflect some generalized cerebral processes regardless of the topography (Hughes et al., 2011; Mizuhara, 2012; Kayser and Tenke, 2015); (vi) topographic analyses revealed that this electrode site is largely representative of the others while capturing the strongest effects (Kayser and Tenke, 2015).

Average power of the raw EMG was calculated over 100 ms according to the usual approach (Merletti, 1999). The resulted value was therefore the area under a power curve, measured in microvolts squared (μV²).

We have followed the conventional NFT protocol for ADHD to decrease TBR (Monastra, 2005). Participants underwent 10 sessions of NFT with eyes open, and each session lasted for 16 min. During training, participants sat in front of the monitor and controlled a kind of computer game by modulating their TBR. As soon as TBR dropped below the threshold, virtual flowers began to grow on the screen. Participants were instructed to find and optimize a strategy/imagery to decrease TBR, which facilitates the flower growth, as follows: “You're going to grow flowers. You can do it by letting yourself feel empty and heavy.” Immediately after the electrodes were attached, we have elaborated on the instruction to provide visual cues for imagery: “It will be easier to relax and let flowers grow up, if you just imagine yourself being a steady, heavy rock. Quiet, steady, and heavy. Relax and let your eyebrows down!” Positive verbal reinforcement was provided in every 2–3 min and the instruction with visual cue was repeated in every 5 min.

In general, thresholds were determined for each participant from 1-min pre-training baseline for each session. Reward criteria were set so that reward thresholds had to be exceeded in 70% of sampled events in a 1,000-ms period, and TBR had to range inhibit above thresholds in 30% of sampled events to receive a reward. For the sNFT group, threshold was calculated based on the mean power in standard frequency bands (4–8 Hz for theta and 12–15 Hz for beta). For the iNFT and the iNFT_EMG groups, threshold was based on mean power of individually adjusted frequency ranges (see section Offline EEG). For the iNFT_EMG group, the mean power of integrated forehead EMG was also taken into account, so that feedback signal (flower growth) appeared only when both TBR and the EMG signal lowered below the thresholds, respectively.

Two-sample T-test showed no significant difference between ADHD subtype groups in psychometrics (|T| ≤ 0.49, p ≥ 0.62) and EMG characteristics (|T| ≤ 0.59, p ≥ 0.92). All subtypes of ADHD exhibited significantly increased reaction time (T ≤ −3.76, p ≤ 0.001) and number of missed stimuli (T ≤ −7.91, p ≤ 0.001) in Go/no-go task, decreased duration (T ≥ 32.95, p ≤ 0.001) in Delayed gratification task and increased forehead muscle tension (T ≤ −4.38, p ≤ 0.001) when compared to healthy peers (Table 1).

The SUBTYPE × CAUDALITY × HEMISPHERE × CONDITION ANOVA showed no significant main effects of SUBTYPE (F = 0.2, p = 0.99, η² = 0.001) and HEMISPHERE [F = 0.1, p = 0.99, η² < 0.001) on alpha indices and TBR. All subtypes of ADHD showed a decreased alpha power in EO condition (i.e., significant main effect of CONDITION: F = 610.74, p < 0.001, η² = 0.215). Factor CAUDALITY showed a significant main effect on alpha1 and alpha2 power, ASM, alpha1 and alpha2 bandwidth, as well as on a1/a2 ratio and TBR (F ≥ 20.54, p ≤ 0.001, η² ≥ 0.012).

The GROUP × CAUDALITY × HEMISPHERE ANOVA showed significant main effect of GROUP on alpha1 and alpha2 power, ASM, alpha1 and alpha2 bandwidth, as well as on a1/a2 ratio and TBR (F ≥ 4.74, p ≤ 0.987, η² ≥ 0,285). HEMISPHERE showed no significant main effect. Table 2 summarizes the post-hoc comparisons of neurophysiological characteristics between the ADHD subgroups and the HC group in a layout similar to Table 1. The Table 2 presents the results of ANOVA of offline EEG as extracted from the PO area.

Table 2 Neurophysiological characteristics of healthy and the ADHD subtype groups. Offline PO EEG and EMG data of healthy and the ADHD subtype groups before the training.

Pearson correlation analysis showed significant relationship between Go/no-Go test results and EEG indices: Both reaction time and number of missed stimuli correlated negatively with individual alpha peak frequency (r ≤ −0.67, p < 0.05), ASM (r ≤ −0.48, p < 0.05), alpha bandwidth (r ≤ −0.40, p < 0.05), and positively with a1/a2 ratio (r ≥ 0.64, p < 0.05), and TBR (r ≥ 0.58, p < 0.05). Duration of the Delayed gratification task correlated positively with individual alpha peak frequency (r = 0.66, p < 0.05), ASM (r = 0.85, p < 0.05), alpha bandwidth (r = 0.60, p < 0.05), and negatively with a1/a2 ratio (r = −0.60, p < 0.05) and TBR (r = −0.74, p < 0.05). Muscle tension measured with EMG also correlated positively with reaction time (r = 0.38, p < 0.05) and missed stimuli number (r = 0.19, p < 0.05) in Go/no-go task, and negatively with the duration in Delayed gratification task (r = −0.40, p < 0.05). According to the multiple linear regression analysis, where pre-training psychological measures of ADHD were included as dependent variables and all EEG and EMG variables were included as predictors, individual alpha peak frequency, individual alpha2 bandwidth, a1/a2 ratio predicted all three psychological measures of ADHD significantly (|beta| ≥ 0.151, |r_semipartial| ≥ 0.081, p ≤ 0.036) (Table 3, typed in bold).

EEG and EMG characteristics are based on the offline EEG and EMG data. At the pre-training session, psychometric, EEG and EMG characteristics showed no difference between the NFT groups (T ≤ 1.55, p ≥ 0.122).

Mixed ANOVA on psychometric outcomes revealed significant main effect of TIME (F = 5.04, p < 0.001, η² = 0.196) and significant TIME × NFT_GROUP interaction on all studied variables (F = 2.26, p < 0.001, η² = 0.116) (Figure 3). Post-hoc tests at the post-training session revealed significant differences between the groups in psychological measures: iNFT and iNFT_EMG groups showed decreased reaction time and number of missed stimuli in Go/no-go task, as well as increased duration of Delayed gratification task in comparison with both sNFT and sham NFT groups (T ≥ 8.84, p ≤ 0.001, η² ≥ 0.266). All these effects remained until the follow-up session 6 months after the NFT in the iNFT_EMG group only (T = 5.30, p ≤ 0.001, η² ≥ 0.182). iNFT group showed improvement only in the number of missed stimuli (Figure 3). NFT efficiency measured with psychological measures was the highest in the iNFT_EMG group (η² > 0.46), while neurofeedback training effect severely diminished in other groups by the time of the follow-up session (Figure 4).

Mixed ANOVA on behavior assessment with Parent and Teacher Rating Scale interview (SNAP-VI) showed significant main effect of NFT_GROUP (F ≥ 4.627, p < 0.001, η² ≥ 0.060) and TIME (F ≥ 2.651, p < 0.01, η² ≥ 0.028) on all variables (Figure 5). Post-hoc tests at the post-training session revealed significant differences between the groups: iNFT and iNFT_EMG groups showed a significant decrease in inattention and impulsivity scores in comparison with both sNFT and sham NFT groups (|T| ≥ 3.239, p ≤ 0.001, η² ≥ 0.060). NFT efficiency was the highest in the iNFT_EMG group when considering inattention (F = 7.703, p < 0.001, η² > 0.060) and impulsivity scores (F = 3.887, p < 0.001, η² > 0.600).

Mixed ANOVA on EEG indices also revealed significant main effect of TIME (F ≥ 4.42, p < 0.001, η² ≥ 0.131) and significant TIME × NFT_GROUP interaction on alpha activity indices and TBR (F ≥ 2.40, p < 0.001, η² ≥ 0.102) (Figure 6). Post-hoc tests at the post-training session revealed significant differences between the groups only in ASM, alpha2 bandwidth, a1/a2 ratio and TBR: iNFT and iNFT_EMG groups showed increased ASM, alpha2 band width, as well as decreased a1/a2 ratio and TBR in comparison with both sNFT and sham NFT groups (T ≥ 5.11, p < 0.001, η² ≥ 0.365). All these effects were present at the follow-up session only for the iNFT_EMG group (T ≥ 3.29, p ≤ 0.001, η² ≥ 0.298).

Mixed ANOVA on muscle tension as measured with EMG showed significant main effect of TIME (F = 4.09, p < 0.05, η² = 0.026) and TIME × NFT_GROUP interaction (F = 7.00, p < 0.001, η² = 0.134) (Figure 6). Post-hoc tests at the post-training session revealed that iNFT_EMG group exhibited muscle tension significantly lower than the other groups (T = 13.40, p < 0.001, η² = 0.308), and this effect were also present at the follow-up session (T = 13.58, p < 0.001, η² = 0.311) (Figure 4).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.