Alpha activity neuromodulation induced by individual alpha-based neurofeedback learning in ecological context: a double-blind randomized study

The neuromodulation induced by neurofeedback training (NFT) remains a matter of debate. Investigating the modulation of brain activity specifically associated with NF requires controlling for multiple factors, such as reward, performance, congruency between task and targeted brain activity. This can be achieved using sham feedback (FB) control condition, equating all aspects of the experiment but the link between brain activity and FB. We aimed at investigating the modulation of individual alpha EEG activity induced by NFT in a double-blind, randomized, sham-controlled study. Forty-eight healthy participants were assigned to either NF (n = 25) or control (n = 23) group and performed alpha upregulation training (over 12 weeks) with a wearable EEG device. Participants of the NF group received FB based on their individual alpha activity. The control group received the auditory FB of participants of the NF group. An increase of alpha activity across training sessions was observed in the NF group only (p < 0.001). This neuromodulation was selective in that there was no evidence for similar effects in the theta (4–8 Hz) and low beta (13–18 Hz) bands. While alpha upregulation was found in the NF group only, psychological outcome variables showed overall increased feeling of control, decreased anxiety level and increased relaxation feeling, without any significant difference between the NF and the control groups. This is interpreted in terms of learning context and placebo effects. Our results pave the way to self-learnt, NF-based neuromodulation with light-weighted, wearable EEG systems.


Materials and methods
Participants. In the NF literature, the common number of included subjects varies from 10 to 20 participants by group 21,[35][36][37]60 . This has been underlined as contributing to overestimated effect size and, by making 'true' effect more difficult to detect, it increases the so-called 'false discovery rate' , that is, the likelihood of having wrongly concluded to a significant effect 61,62 . Here, we included forty-eight healthy volunteers, divided in two groups of 25 and 23 participants respectively (see below; mean age: 33 Table S1 for more details). While this limited our sensitivity to effect sizes of at least 0.028-0.048 in eta-squared (Cohen's f = 0.17-0.22) at 0.80 statistical power (as computed with G*Power 3.1.9.2, 'computation of sensitivity for repeated-measure ANOVA' , with type 1 error rate alpha = 0.05, correlation among repeated measures = 0.5, non-sphericity correction ε = 1 to 0.5, and a 12 within-subject repeated measures design), it was based on the literature and resources constraints 63 implying a follow-up across 12 weeks for each subject (see the "Experimental protocol" section). All participants declared having normal or corrected-to-normal vision, no hearing problem, no history of neurological or psychiatric disorders, no ongoing psychotropic drug treatment and no or little NF or BCI experience. Participants were blindly assigned either to the NF group-who received real NF-or to the control group-who received sham-FB. For the purpose of the sham-FB design construction, the first N participants were assigned to the NF group. Only the experimenters and the data analysts knew the existence of the two groups and that the first N subject(s) was/were in the NF group. However, the experimenters and the data analysts were blind to N and blind to the random assignment after N. This resulted in a double-blind sham-controlled design with 25 subjects in the NF group and 23 in the control group. The blind assignment was maintained until the end of the experiment. No test was done to know if the participants suspected the existence of two groups and their assignment to one of these groups. Participants were enrolled from the general population through advertisements in science and medical schools in Paris, through an information mailing list (RISC, https:// expes cienc es. risc. cnrs. fr) and through flyers distributed in companies in Paris (France).
Participants completed the protocol in three different locations: at the Center for NeuroImaging Research (CENIR) of the Paris Brain Institute (ICM, Paris, France) (N = 20 participants, NF group: 10, control group: 10), at their workplace 14 (N = 14, NF group: 8, control group: 6), or at home (N = 14, NF group: 7, control group: 7). The 20 participants who performed the protocol at the CENIR were part of those planned in the study approved by French ethical committee (CPP Sud-Ouest et Outre Mer I, ref. 2017-A02786-47), registered on ClinicalTrials. gov (ref. NCT04545359), although the present study was not part of this clinical study. For these participants, a financial compensation was provided at the end of the study for the time taken to come to the lab. The 28 other participants followed the same protocol but performed it in a real-life context (at work, at home). Moreover, all participants gave written and informed consent in accordance with the Declaration of Helsinki.
EEG recording and preprocessing. Brain activity was recorded by two gold-coated dry electrodes placed on parietal regions (P3 and P4) (Melomind, myBrain Technologies, Paris, France; Fig. 1). Ground and reference were silver fabric electrodes, placed on the left and right headphones respectively, in mastoid regions.
EEG signals were amplified and digitized at a sampling rate of 250 Hz, band-pass filtered from 2 to 30 Hz in real-time and sent to the mobile device by embedded electronics in the headset. The headset communicated via Bluetooth with a mobile NF application, which processed EEG data every second to give the user auditory feedback about his/her alpha-band activity (see below). A DC offset removal was applied on each second of data for each channel and a notch filter centered at 50 Hz was applied to remove the powerline noise. Real-time   64 for more details). In Grosselin et al. 64 , we showed that this algorithm has an accuracy higher than 90% for all the studied databases. This algorithm was used to detect noisy segments (LOWq) which were excluded from posterior analysis.
Experimental protocol. Based on previous studies 15, 65 , we proposed a protocol consisting in 12 NFT sessions, with one session per week (Fig. 2). Each session was composed of 7 exercises of 3 min (total: 21 min), which corresponded to 4.2 h of training. At the beginning and end of each session, two-minute resting state recordings were performed and the participant completed the Spielberger State-Trait Anxiety Inventory (STAI, Y-A form, in French 66 )-to assess his/her anxiety state level-and a 10-cm visual analog scale (VAS) indicating his/her subjective relaxation level (relax-VAS). These resting state recordings were not analyzed here as they are out of the scope of this study focused on neuromodulation. Moreover, at the end of each 3-min exercise, the participant indicated his/her subjective level of feedback control on a 10-cm VAS (control-VAS)-the left side indicating the feeling of no control; the right bound indicating a feeling of perfect control.
Neurofeedback training procedure. The NF paradigm targeted alpha rhythm centered on the individual alpha frequency (IAF). Before each NFT session, a 30-s calibration phase allowed computing IAF using an optimized, robust estimator dedicated to real-time IAF estimation based on spectral correction and prior IAF estimations 68 . More precisely, the spectrum was corrected by removing the 1/f trend estimated by an iterative curve-fitting procedure. Then, local maxima were detected in the corrected spectrum between 6 and 13 Hz as the downward going zero crossing points in the first derivative of the corrected spectrum. If the presence of an alpha peak was ambiguous, the algorithm selects the most probable one based on the IAF detected in previous time windows. See Grosselin et al. for more details 68 . All participants were instructed at the protocol explanation to close their eyes during the recordings. This instruction was reminded audibly at the beginning of each calibration. They were also instructed to be relaxed and try to reduce the auditory feedback volume throughout the exercises of different sessions. Previous research showed that providing no strategies yielded to better NF effects 57 . Here, the participants were aware that the feedback volume would decrease with relaxation, but no explicit strategies were provided to them as such to allow them to reduce the auditory feedback volume; they were told to try their own strategies, which we report in the Supplementary Material as advised in the CRED-nf checklist 17 . A relaxing landscape background (e.g. forest sounds) was played at a constant, comfortable volume during each exercise. The audio feedback was an electronic chord sound added to this background with a volume intensity derived from EEG signals. More precisely, the individual alpha amplitude was computed in consecutive 1-s epochs as the root mean square (RMS) of EEG activity filtered in IAF ± 1 Hz band (NF index); it was normalized to the calibration baseline activity to obtain a 0-1 scale, which was used to modulate the intensity of the feedback sound (V) in the NF group. More precisely, for each session, a baseline value was obtained from alpha activity during the corresponding 30-s calibration phase without the low quality EEG segments as assessed by a dedicated algorithm (see "EEG recording and preprocessing" section above). Coefficients were applied to this baseline value in order to define the lower (m) and upper (M) thresholds of alpha activity during the session. During the NFT, V was varied as a reverse linear function of the individual alpha amplitude relative to these upper and lower bounds. If the individual alpha amplitude was becoming lower than m, then V was set to 1 (maximal). If an alpha amplitude beyond M was reached, then V was set to 0 (minimal). For the EEG segments detected as noisy (LOWq quality) during the www.nature.com/scientificreports/ preprocessing step, V was set to 1. For the participants in the control group, the instruction was identical but they received sham-FB, which was the replayed feedback variations from another subject randomly chosen from the NF group at the same training level (i.e. session). For instance, a participant in the control group at the 3rd session received the auditory feedback generated and received by a random subject from the NF group at the 3rd session.
Data analysis. NF index and learning score. For each participant and each training session, we first computed the average value of the NF index (before normalization) for every exercise. Second, in order to take into account inter-subject variability at the first session for NF index (see Fig. 3a and Supplementary Fig. S13), we built an NF learning score (ΔD(t))-from the NF index variations across exercises and sessions 69 . To do this, we computed the median value (med) of the NF index across the 7 exercises of the first session; then, for each session t, we computed D(t), the number of NF index values (1 by second) above or equal to this median value med. This cumulative duration was divided by the total duration of the training session cleaned from LOWq segments (maximum 21 min) in order to express D(t) by minute, and transformed into percent change relative to the first session, as follows (Eq. (1)): Theta and low beta activities. To study the selectivity of the neuromodulation only for the targeted alpha activity, we analyzed the between-session evolutions of theta (4-7 Hz) and low beta (13-18 Hz) activities, as control outcomes 69 . For each subject, on each exercise and session, theta activity was computed every second as the RMS of EEG activity filtered between 4 and 7 Hz in 4-s sliding windows, on epochs with high or medium quality (see 64 for details about signal quality computation). We then averaged these RMS values for each session. Similar computations were performed for the EEG activity between 13 and 18 Hz (low beta activity).
Signal quality. As encouraged in 17 , the quality of EEG signals was analyzed to assess the poor quality EEG data prevalence between groups and across sessions. For each participant, session, and exercise, the quality of each 1-s EEG epoch recorded by each electrode was determined by a classification-based approach according to three labels: HIGHq, MEDq, and LOWq (see 64 for more details). A quality index Q was then computed for each electrode, during each exercise, as in Eq. (2): www.nature.com/scientificreports/ with: #HIGHq, #MEDq, #LOWq indicating the number of high, median, low quality epochs and N, the total number of quality labels during the session. Finally, the average value of Q was computed from the two electrodes for each exercise.
Self-report outcomes. The raw scores of the STAI-Y-A (between 20 and 80) and relax-VAS (between 0 and 10) were computed pre-and post-session. The subjective level of feedback control was measured within-and between-session on the control-VAS (between 0 and 10). The raw scores of the STAY-Y-B, PANAS and PSS were obtained pre-and post-program; these latter outcomes are reported in Supplementary Material.
Statistical analyses. All statistical analyses were performed using R (v.4.0.2; R Core Team, 2020) and lme4 package 70 . We used Linear Mixed Models (LMMs) 71,72 , because LMMs allow handling missing data, variability in effect sizes across participants, and unbalanced designs 73 . Available data in this study are detailed in Supplementary Table S8.
For all LMM analyses, the NF group at the first session was set as the level of reference in order to specifically estimate the effects of NFT in this group. For each outcome variable studied, the choice of the random factors was done comparing the goodness of fit of the models that converged with different random factors, based on Akaike Information Criterion (AIC) 74 , Bayesian Information Criterion (BIC), log-likelihood comparison (logLik) and by running an analysis of variance (anova) between models. The detailed procedure for each outcome variable can be found in Supplementary Material in Sect. 6. To be concise in the main text, the random factors chosen were directly reported between parenthesis in the LMM equations below.
Similarly to 75 , to analyze the within-and between-session NFT effects on the NF index we used fixed effects of session, exercise, group, and the 2-way interactions between session and group and between exercise and group in the following equation (Eq. (3) as coded in R, with a colon indicated an interaction between terms): Results (see "5. NF index and feeling of control across exercises: U-curves" section in Supplementary Material) indicated that the effect of exercises followed a U-curve. Therefore, the exercises were coded as a quadratic term, that is, exercises 1 to 7 were coded as 9, 4, 1, 0, 1, 4, and 9. The sessions were coded as a numeric variable between 0 and 11. Equation (3) was also used for the analysis of the control-VAS scores with 1 + session|subject_id as random effects structure.
For the analysis of NF learning score and the signal quality index, we used the following LMM equation (Eq. (4)): Equation (4) was also used for the analysis of theta and low beta activities with only a random intercept by participant (1|subject_id).
For the STAI-Y-A outcome, we used LMM with session, phase (pre-or post-session), group, and the 2-way interactions between session and group and between phase and group as fixed effects (Eq. (5)). This model was also used for the analysis of relax-VAS with 1 + phase|subject_id as random effects structure.
For each model, parameter β for the effects of interest were estimated by fitting the models on the corresponding dependent variable, using the Restricted Maximum Likelihood (REML) approach. P-values were estimated via type III Analysis of Variance on the LMM with Satterthwaite's method, using the anova() function of the lmerTest package of R 76 . For all variables of interest, we set p < 0.05 as statistically significant. When there was an interaction between a factor of interest and group, LMM models were fitted in each group separately, with session and exercise or phase-as adequate-as fixed-effect factors. For these latter analyses, we used a random intercept by participant because more complex model structure generally failed to converge for at least one group 77 . All the results of these LMM and anova analyses are reported in Supplementary Material.
Moreover, for the analysis of session effects in the theta and low beta bands, we conducted supplementary equivalence tests using the TOST procedure 78 , to examine (i) the equivalence of the session effect in the NF and control groups for the theta activity and (ii) if we could conclude to an absence of change across sessions for the low beta activity. The results of these tests are presented in Supplementary Material Sect. 14.
Additionally, to check the variability between groups at the first session, we performed, for each outcome variable of interest, an independent t-test between groups. The results of these t-tests are presented in Supplementary Table S63.
Correlation analyses were also performed between NF index and self-report outcomes in each group as detailed in Sect. 13 of Supplementary Material.
The analyses were not pre-registered. The primary outcome measures in this study were the NF index, the STAI-Y-A and relax-VAS outcomes. The NF learning score, the theta and low beta activities, the signal quality index, and the subjective feeling of control (control-VAS) were secondary outcome measures. All other analyses were additional analyses performed on reviewers' requests.  Tables S29 and S30). These findings indicated an increase of the NF index across sessions, specific to the NF group (Fig. 3a). Although these results could be due to a baseline difference at the first session, an independent t-test between the mean levels of NF index of each group at session 1 did not show a significant difference between groups (t(46) = − 1.8, p = 0.0789). See "11. Group comparison at the first session" supplementary section and supplementary Table S63 for details. In addition, to normalize changes across sessions relative to the NF index at the first session, we built an NF learning score. This allowed us to analyze the progression of the trained activity across sessions taking into account the activity at the first session. The analysis of the NF learning score did not show a significant interaction between session and group (F(1, 46.325) = 3.27, p = 0.077) (see Supplementary Tables S31 and S32). However, following on our a priori hypothesis, we looked at the session effect in each group (Supplementary  Tables S33, S43, S35 and S36). The analysis of the NF learning score in each group confirmed a specific NF-based neuromodulation. Indeed, our analyses showed a significant effect of session only for the NF group (β = 1.14, CI [0.20, 2.08], F(1, 272.19) = 5.67, p = 0.018) (see Supplementary Tables S33 and S34), which indicates that the NF learning score increased across sessions in the NF group (Fig. 3b). The effect of sessions was not significant for the control group (Supplementary Tables S35 and S36). Additional individual linear regressions of the NF learning score (see Supplementary Fig. S8 and S9) showed that 80% (20/25) of the participants from the NF group had a positive regression slope across the 12 sessions, while the slope was positive for 48% (11/23) of the participants from the control group.
In addition, there was a significant effect of exercise on the NF index, reflecting the quadratic pattern of the NF index across exercises (F(1, 45.940) = 26.55, p < 0.001, see Supplementary Tables S2 and S3). The non-significant www.nature.com/scientificreports/ interaction between exercise and group indicated that this effect did not statistically differ between groups (F(1, 45.940) = 1.76e−03, p = 0.967) (Fig. 4).

Selectivity of the neuromodulation on alpha activity.
To investigate the selectivity of the neuromodulation relative to the targeted alpha activity, we checked if some specific neuromodulation for the NF relative to the control group occurred for EEG frequency bands close to the alpha band. Thus, we analyzed EEG activity in the theta (4-7 Hz) and low beta (13-18 Hz) bands. For theta activity, there was a significant effect of session (F(1, 523.07) = 6.50, p = 0.011), without any statistically significant interaction between group and session (F(1, 523.07) = 2.11, p = 0.147). This reflected an overall increase of theta activity with no statistically significant difference between the NF and the control groups. As absence of evidence is not evidence of absence, we further tested if the session effect in the NF and in the control group was equivalent (using the TOST procedure 78 ). This did not allow demonstrating statistical equivalence. Therefore, the only reliable effect for theta activity was the main effect of sessions. See Supplementary Tables S37, S38, S68 and Supplementary Fig. S10 for details. For low beta activity, the main effect of session was not significant (F(1,523.04) = 0.15, p = 0.694) and the interaction between session and group was not significant either (F(1, 523.04) = 3.70, p = 0.055). However, considering that the p value of this interaction could be deemed as 'close to significance' , we further checked if the session effect was significant in either group. The analyses in each group did not reveal any significant session effect either in the NF group  Supplementary Tables S45 and S46 and Supplementary Fig. S11). Moreover, a chisquare analysis (see "12. Study of LOWq, MEDq and HIGHq proportions" section in Supplementary Material) did not show any statistically significant difference in the proportions of LOWq, MEDq and HIGHq EEG segments between the first and last sessions (X 2 = 0.009, df = 2, p = 0.9955).

Self-report outcomes. Relaxation and anxiety levels. STAI-Y-A.
The state anxiety level decreased significantly from pre-to post-session (phase effect: F(1, 46.137) = 24.77, p < 0.001). The interaction between phase and group was not significant (F(1, 46.137) = 2.18, p = 0.147) (Fig. 5a). Moreover, although the overall mean of STAI-Y-A scores decreased across sessions (Fig. 5b), the session effect on STAI-Y-A scores was not

Correlations between NF index and self-report outcomes.
We examined these correlations at each session and the correlation between the slope of NF index and the slope of self-report outcomes-in terms of relaxation, anxiety, and feeling of control-across sessions (Supplementary Tables S65, S66 and S67). A few significant correlations were found at some sessions, but none of these was significant after correction for multiple comparisons. There was no significant correlation either between the slopes of NF index and self-report outcomes.

Discussion
In this study, we proposed a double-blind sham-controlled randomized study of the neuromodulation induced by individual alpha-based NFT over 12 weekly sessions using a strictly controlled sham-FB condition as control, in healthy adult participants. NFT was performed with a wearable, dry sensors headset, which delivered intensity-modulated auditory feedback based on EEG signal amplitude in individual alpha frequency band. To avoid non-contingency between produced efforts and the resulting feedback evolution for the control group 22 , the control condition consisted in delivering sham-FB-a feedback replayed from randomly chosen users of the NF group at the same training stage. Hence, all participants benefited from the proposed NFT experience, but only those of the NF group experienced a link between the feedback and their own alpha activity. In addition, all the participants performed the task immersed in a relaxing auditory landscape, with their eyes closed for 21 min, thus constituting a common reinforcer context for relaxation.
First of all, we wanted to assess the NF learning of individual alpha-band activity upregulation. NF learning refers to the capacity to self-regulate a targeted activity in the desired direction across training sessions 34,41,42,57,79 . www.nature.com/scientificreports/ More specifically, we hypothesized a neuromodulation specific to the NF training, that is to say, only the NF group was expected to increase individual alpha activity across training sessions 80 . Even if the averaged values of NF index were similar at the 12th session in both groups, our analyses of the NF index and the NF learning score confirm a specific session effect in the NF group, with significant linear increase across sessions in this group only. This finding demonstrates, across the training program, a specific neuromodulation induced by the link between individual alpha activity and FB. Indeed, the use of a randomized double-blind protocol together with the strict sham-FB control condition in a reinforcer context allowed us to control for different confounding factors which may contribute to NFT effects. In particular, it allowed controlling for context, task, reward, and performance, avoiding potential motivational biases in NF versus control conditions 22 . One may wonder if training another frequency band could have constituted an alternative sham condition for the control group. However, as mentioned in "Introduction" section, such control condition may induce incongruity between task instruction and the target activity in the control condition. This could render the task more difficult or less rewarding for the subject in the control condition, due to this incongruity. This is why we chose the present yoked feedback. We observe that the NF index seemed different between NF and control groups at the first session. This could be explained by the inter-subject variability of alpha rhythm 81 . We tested this difference as well as that of other outcome variables in the first session; it was not significant (cf. Supplementary Table S63). Similarly, one may note that NF index values seemed similar between groups at the 12th, final session. However, the important inter-individual variability in alpha activity makes it important to consider across sessions effects, as we did in our analyses, rather than NF index value at either first or final session.
In this study, we also examined if the neuromodulation was selective of the targeted activity [82][83][84] . As proposed in 69 , we analyzed two adjacent frequency bands (theta and low beta). We found an overall increase across sessions for the theta band and no significant change for the low beta band. This indicated that the neuromodulation was selective of the alpha activity insofar as there was an absence of evidence for similar effects in the theta and beta bands. To the best of our knowledge, this is the first evidence of selective longitudinal alpha-band neuromodulation (over 12 weeks) in a double-blind randomized study implying healthy participants trained with a wearable dry sensors NF device.
Interestingly, we could not assess an alpha neuromodulation within sessions (across exercises) as one may have expected it 34,42,79 . In fact, we observed a U-shape pattern for the dynamics of alpha band activity during NF learning across exercises. To observe the NF learning effects, multiple sessions are required 41,85 , in order for the participants to find their own strategies to succeed in the task 57 . In contrast, the effects observed withinsession may not only be related to relaxation training but also to other processes put at play during each session. Alpha activity is a spontaneous but complex rhythm associated with several cognitive states and processes. Its modulation has been predominantly related to vigilance, attention 86,87 , awake but relaxed state 50,55,56,[88][89][90][91][92] . The alpha activity change across exercises during the sessions could reflect the different cognitive processes involved by the task: continuous monitoring of the feedback may have required heightened focused attention 93,94 , error www.nature.com/scientificreports/ detection 95,96 , and working memory processes 97 during the first training minutes, allowing participants to progressively adapt their cognitive strategy and mental state to the task. It is important to note that the within-session U-shape pattern of alpha activity was observed in both the NF and control groups. This supports the idea that the sham-FB condition allowed us to rigorously control for the task performed by the subjects. Altogether, the specific neuromodulation of alpha activity induced by NFT was revealed only in the longitudinal effect across the twelve sessions. We also examined the possible effect of EEG data quality on NF learning 17 . EEG quality did not change significantly across sessions in either group. One may wonder if we monitored the compliance of keeping eyes closed during the recordings because of the potential effects of eyes open and eyes closed on alpha activity. Even though this instruction was reminded auditorily at the beginning of each calibration, it could not be checked for the 28 participants who underwent the NFT sessions at home or at work. It has to be noted that if the participants didn't respect this instruction during the calibration or the training sessions, this may have had an impact on data quality, hence on the feedback. For future experiments, it will be interesting to find a way to monitor this aspect of the task.
In this study, we were also interested to know if self-report benefits would be induced by the NFT and if a difference would be found between groups knowing the common reinforcer (relaxing) context of the protocol. We investigated the self-report changes in terms of relaxation and anxiety levels pre-and post-session and across the training program. We found significant benefits in terms of relaxation and anxiety from pre-to post-session, as well as a slight reduction of anxiety level and a significant increase of relaxation across sessions, but without any statistically significant group difference. This finding is reminiscent of Schabus et al. 98 , who performed a wellcontrolled double blind NF study targeting sensorimotor rhythm in insomnia with NF and sham groups. They found some specific neurophysiological effect of NF (relative to sham) condition but non-specific self-report, psychological effects in both NF and sham conditions. In the present study, the non-specific self-report benefits of NFT may be explained by the use of sham-FB condition and the NF task proposed in our protocol, which could produce the same immersive, relaxing experience in the participants of both the control group and the NF group.
There was no significant correlation between NF index and self-report outcomes either. Thus, the self-reported benefits were not found to be specific to the NF operant learning. While this absence of evidence is not a proof of the absence of any specific effect, we propose that self-reported benefits in our study may be explained by nonspecific mechanisms of the NFT, such as the psychosocial factors (like education level, locus of control in dealing with technology, capacity to be mindful, field of work, etc.) 30,31,33 , relaxing training context, the instructions (closed eyes during a break of 21 min), and repetition-related effect 17 , in line with the view that placebo effect can play a role in psychological outcome of neurofeedback 30,59 . Note that education levels and the professions of the participants had almost the same repartition in both groups as well as the frequency of practice of meditation, sophrology, relaxation, arts (see Supplementary Tables S2, S3, S4, S5, S6 and S7). Furthermore, we must notice that all the subjects involved in our study were ranked as low to moderately anxious, which might have contributed to the lack of difference between groups. Indeed, Hardt and Kamiya 99 , in their alpha-upregulation NFT study, observed reduction of anxiety level for high but not low anxious subjects. Further investigations with high anxious or clinical participants should allow to test if benefits in terms of relaxation and anxiety may be highlighted specifically for the NF group.
Overall, our findings showed that NFT induced positive self-report benefits for all participants, without any evidence for a significant link between these self-report benefits and the alpha activity modulation specifically induced in the NF (relative to the control) group. Indeed, the links between self-report outcomes and neurophysiology are complex and include several factors 17 , such as cognition, attention, motivation 33 , training frequency 85 , but also the choice of the neuromarker itself 34,58 . In this study, we chose, as in most NF protocols aiming at anxiety reduction, stress-management or well-being 14,15,55,56 , to use alpha activity as a biomarker for its known link with relaxed or meditative states 35,[43][44][45][46][47][48][49] . However, the alpha activity is not the unique biomarker of stress management, anxiety, relaxation and well-being. For instance, it can be a marker for attention 93,94 or memory 97 . In addition, other biomarkers such as theta activity [100][101][102] , beta activity 103,104 or the ratio theta/alpha 43,45 have also been associated with stress and/or anxiety reduction. Such biomarkers could be interesting targets to investigate in order to optimize our NF protocol. Further investigations should focus on the research of specific biomarkers related to psychophysiological factors, for example using neurophenomenology to study the link between neural activity modulation and participant's inner experience 105 .
Finally, to study the effect of sham-FB, we asked participants to assess their feeling of control during the training 106 . We found an increase of the feeling of control across sessions in both groups, which suggests that participants of the control group were not aware of the non-contingency between their efforts and the feedback signal and had a qualitatively similar experience as those of the NF group. Although the increase in the feeling of control across sessions seemed more marked in the NF group, there was no significant difference between groups on this outcome variable. This emphasizes the closely controlled nature of our sham control condition. It suggests that our manipulation of the sham feedback remained fully implicit to the subjects. One may note that we did not check the locus of control of the participants in dealing with technology, which may have an impact on the training 33 .
To conclude, our study demonstrated an upregulation of the individual alpha-band activity specific to the NF group with a wearable dry-sensor EEG device across multiple sessions of NF training. In contrast, self-reported effects in terms of relaxation and anxiety were observed in both the NF and the control groups. Even if the relationship between the targeted EEG modulation and self-report outcomes is complex and remains to be fully elucidated, this study with a wearable dry-sensor EEG device underlined that NF can be used outside the lab to investigate and generalize NF learning mechanisms in ecological context.

Data availability
The datasets generated during the current study are not publicly available due to the subject's consents and restrictions of the ethics protocol to protect the privacy of subjects involved in the study.