Neurofeedback and attention modulate somatosensory alpha oscillations but not pain perception

To determine the maximal sample size for a sequential Bayes factor design with maximal N (SBF+Nmax) and to evaluate the properties of such a design, we conducted a Bayes factor design analysis [37,41]. To this end, we first computed an open-ended SBF design for one-sided paired t tests under both H1 and H0, yielding a distribution of sample sizes at the stopping point for both hypotheses (parameters: Cohen’s d H1 = 0.41, Cohen’s d H0 = 0, Cauchy prior distribution with scale parameter r = √(2)/2 and truncation at zero, directionality H1 = “greater”, minimal sample size = 20, step size = 5, maximal sample size = 300, lower BF10 boundary = 1/10, upper BF10 boundary = 10, number of simulated studies = 10,000. Note that the maximal sample size for simulations was set to 300 to explore the development of power across a wide range of sample sizes). Subsequently, we determined the sample size at which 90% of studies reached the defined BF10 boundaries (1/10 or 10) under the H1. The resulting sample size was then used to evaluate a corresponding SBF+N_max design with respect to false positives and false negatives rates.

Fig A in S1 Text visualizes the results of an SBF+N_max design with the obtained N_max = 95. Under H0, 86.0% of simulated studies correctly indicated evidence for the absence of an effect (58.3% at BF₁₀ boundary = 1/10 + 27.7% at the BF₁₀ boundary = 1/3), while the false positive rate was acceptably low (2.3%). The average sample size at the stopping point was 68. Under H1, 95.6% of simulated studies correctly indicated evidence for the presence of an effect (90.9% at BF₁₀ boundary = 10 + 4.7% at the BF₁₀ boundary = 3), while the false negative rate was acceptably low (0.3%). The average sample size at the stopping point was 49. Thus, the chances of detecting an effect, if it exists, were rated as sufficiently high for the proposed SBF+N_max = 95 design.

Since effect size estimates (ES) are subject to uncertainty, we explored the effect of our assumptions by conducting BFDAs for further ES. The performance of a sequential Bayes Factor design with different ES and sample sizes is illustrated in Fig B in S1 Text.

Methodology and terminology of this study are based on recent recommendations for EEG and neurofeedback studies [17,42,43]. Our EEG-based neurofeedback study employs a within-subjects, bidirectional control design with 2 verum and 2 sham conditions which were completed during 2 sessions (Fig 2). The design was adapted from recent neurofeedback studies which successfully modulated the asymmetry of alpha oscillations [30,31]. In a first verum neurofeedback condition, participants were instructed to focus attention on their right hand and the up-regulation of alpha oscillations in the right hemisphere relative to alpha oscillations in the left hemisphere was incentivized through neurofeedback (attention right training, ART_NF). In a second verum neurofeedback condition, participants were instructed to focus attention on their left hand and the down-regulation of right relative to left alpha oscillations was incentivized (attention left training, ALT_NF). To control for nonspecific effects of the neurofeedback procedure such as expectation effects evoked by the explicit instruction, the 2 verum conditions were accompanied by 2 matched sham neurofeedback conditions administered in a separate session. During the sham neurofeedback conditions, participants obtained the same instructions, i.e., focus attention on your right hand (sham attention right training, ART_sham) or focus attention on your left hand (sham attention left training, ALT_sham). However, the feedback signal did not mirror their brain activity. Instead, the feedback signal and the corresponding reward of the last matching verum condition completed by a previous participant, i.e., ART_NF for ART_sham and ALT_NF for ALT_sham, were replayed (yoked feedback). Thus, on a group level, feedback signal and compensation were identical between verum and sham conditions supporting blinding of participants and experimenters. To implement sham conditions for the first participant(s), data from a “participant zero” was collected, who only completed the verum trainings and was excluded from the analyses. To avoid order effects, sessions were separated by 7 days at least and the order of conditions between (verum versus sham) and within sessions (ART versus ALT) was pseudo-randomized.

Fig 2. Experimental procedure.

(a) Attention-based neurofeedback training. During neurofeedback, the somatosensory AAI was calculated in real-time based on 1,000 ms EEG data segments and was fed back to participants using the visibility of neutral face images [44]. This feedback cycle was updated every 100 ms. To control for nonspecific effects, the experiment entailed 2 verum neurofeedback conditions during which alpha asymmetry was regulated in opposite directions as well as 2 sham neurofeedback conditions with identical instructions but yoked feedback. During the attention right training (ART_NF), participants were instructed to focus attention on their right hand and the up-regulation of right relative to left alpha oscillations was incentivized. During the attention left training (ALT_NF), participants were instructed to focus attention on their left hand and the down-regulation of right relative to left alpha oscillations was incentivized. (b) Real-time visual feedback. AAI values and corresponding feedback signals for the verum conditions are displayed. For ART, AAI values above −0.6 led to an increase of image visibility until full visibility was reached at an AAI of 0.6. For ALT the relationship between AAI and image visibility was reversed. In addition, a small, central fixation cross was superimposed on the images to support image fixation. (c) Experimental design. Verum and sham neurofeedback conditions were completed in a pseudo-randomized order during 2 separate sessions. Each session entailed verum or sham conditions (ALT_NF and ART_NF or ALT_sham and ART_sham, 40 trials each) as well as baseline assessments of pain sensitivity (pain_pre, 20 trials) and brain activity (EEG_pre, 5 min resting state with eyes open). The latter, baseline measures of pain sensitivity and brain activity, were obtained to evaluate predictors of regulatory success and will be analyzed and reported in a separate manuscript. The sequence of events of neurofeedback trials is shown on the right. After a fixation period of 3 s, the regulation period of 15 s began. Immediately afterwards, a noxious stimulus was applied. To avoid that pain-related brain responses were confounded by visual offset responses, the last feedback signal of the regulation period remained on the screen for another second before turning into a fixation cross. Two seconds later, an auditory and visual cue prompted participants to rate the perceived pain intensity. Finally, the financial reward earned on a given trial was displayed. The face stimulus shown as an example for the visual feedback is derived from [44], image ID: FNES. AAI, alpha asymmetry index; ALT, attention left training; ART, attention right training; EEG, electroencephalography; NF, neurofeedback.

https://doi.org/10.1371/journal.pbio.3002972.g002

Both sessions comprised a fixed sequence of events aimed at capturing changes of alpha asymmetry, pain perception, and pain processing that occur during the neurofeedback training (Fig 2). Each experimental condition entailed 40 trials with 15 s of neurofeedback training, which were followed by a brief noxious laser stimulus applied to the dorsum of the left hand. Three seconds after the noxious stimulus, participants verbally rated the perceived pain intensity on a numerical rating scale (NRS) ranging from 0 (“no pain”) to 100 (“worst tolerable pain”). To enhance motivation, participants received feedback regarding their performance-based financial compensation at the end of each trial. Thus, the total neurofeedback training time per condition was 10 min. Alpha asymmetry during these training runs served as read-out of regulation success. Pain ratings served as read-out for pain perception. Brain responses to noxious stimuli served as read-out for pain processing. To quantify baseline measures, pain perception, pain processing, and resting-state brain activity were assessed before (pain_pre and EEG_pre, respectively) the neurofeedback training. During the pain_pre assessment, participants rated 20 noxious laser stimuli applied to the left hand on the same NRS used during neurofeedback and pain ratings served as read-out for pain perception. Brain responses to noxious stimuli served as read-out for pain processing. The EEG_pre assessment entailed 5 min of resting-state recordings with eyes open during which a central fixation cross was displayed and served as read-out for the brain state during rest.

Noxious stimuli were applied using cutaneous laser stimulation, which enables the well-controlled stimulation of nociceptive pathways without concomitant stimulation of tactile pathways [45]. All stimuli were applied to the dorsum of the left hand using a neodymium:yttrium-aluminum-perovskite (Nd:YAP) laser (Stimul 1340, DEKA M.E.L.A. srl, Calenzano, Italy) and the following settings. Stimulus duration was set to 4 ms and stimulus diameter to 7 mm. Laser intensity was set to 3.5 J, which induces stable brain responses while being well tolerated [46]. To avoid tissue damage and minimize habituation/sensitization effects, stimulation sites were changed slightly after each stimulus. For safety reasons, both study personal and participants wore safety goggles throughout the experiment.

Motivated by recent studies [30,31,40], we designed a short-term neurofeedback protocol with real-time data analysis and feedback visualization using MATLAB (version: R2020b, Mathworks, Natick, Massachusetts, United States of America) and the MATLAB-based toolboxes FieldTrip (version: 20210212, [47]) and Psychtoolbox-3 (version: 3.0.17 beta, https://psychtoolbox.org). To this end, EEG data was streamed from an acquisition computer (operating system: Windows 10) to a second computer (operating system: Ubuntu 20.04.2 LTS) responsible for data processing and stimulus presentation in real-time and is stored in a buffer. Every 100 ms, 1,000 ms segments were accessed from this buffer and analyzed as follows. Data were demeaned, and power estimates from 8 to 48]. Resulting power estimates were then used to calculate the alpha asymmetry index (AAI), defined as:

where α_rS1(t) and α_lS1(t) represent alpha power in time window t in right and left somatosensory regions, respectively. AAI values range from −1 to 1 when power is purely left- or right-lateralized, respectively, and were communicated to participants by altering the visibility of neutral face images (see paragraph below and section “Image visibility” in the Supporting information for details).

Participants were instructed to use spatial attention towards the left or right hand to enhance the visibility of face images as much and as long as possible until the pain stimulus is applied [30]. To enhance motivation, participants received a performance-based financial compensation for each trial. This compensation reflected the average AAI on a given trial and amounted to up to 0.25 € per trial (ART: linear increase from 0–0.25 € for positive AAI values with a bonus of 0.25 if the AAI was larger than 0.6; ALT: linear increase from 0–0.25 € for negative AAI values with a bonus of 0.25 if the AAI was smaller than −0.6).

To create a motivating feedback visualization while avoiding potential effects of emotion or gender, AAI values were fed back to participants by altering the visibility of neutral face images presented centrally on a screen. Specifically, 8 neutral face images depicting an average female and male face from 4 different angles were taken from the Averaged Karolinska Directed Emotional Faces (AKDEF) data set [44] (image IDs: F/MNEFL, F/MNEFR, F/MNEFHL, F/MNEHR; see Fig C in S1 Text for a visualization). For each trial, one of these images was randomly chosen, thus avoiding systematic effects linked to one of the images. Depending on the training condition, image visibility was altered in opposite directions. During ART_NF, AAI values above −0.6 led to an increase of image visibility until full visibility was reached at an AAI of 0.6. During ALT_NF, AAI values below 0.6 led to an increase of image visibility until full visibility was reached at an AAI of −0.6 (see below section “Image visibility” and Fig C in S1 Text for details). During sham conditions, the feedback signal of the last verum condition completed by another participant was replayed (yoked feedback) irrespective of the current alpha asymmetry.

Let c be the parameter controlling visibility. It ranges from zero (no visibility) to one (full visibility) and can be defined as a function of the AAI and the condition ALT/ART as follows:

To achieve different levels of visibility, a scrambled version of each image was created by randomly perturbing the phase angle of the 2D Fourier transformation of the original image. The magnitude of the perturbation is proportional to the visibility parameter c introduced above. In this manner, low-level image properties such as spatial power spectrum and luminance are held constant across varying AAI-values within a trial. The procedure of generating the scrambled images is explained in the following. Let be the matrix representation of a gray scale image with n_x by n_y pixels. Let 〈I〉 denote the mean value of the image. The 2D FFT of the demeaned image is given by

The scrambled image is obtained as the inverse 2D FFT of the Hadamard product of and a complex-valued perturbation matrix P(c):

The entry in the k-th row and l-th column of P(c) is

where Δϕ_kl are random phase angles drawn from a uniform distribution on the interval [−π, π]. To ensure that the inverse 2D FFT is real-valued, the random phase angles satisfy the symmetry property . In the scrambling of the AKDEF images used in this study, the lowest frequency components of the image were excluded from the random phase perturbation. This was necessary since the scrambled images would otherwise show a pattern consisting of horizontal and vertical stripes.

To avoid sudden jumps in the feedback signal arising from rather fast fluctuations of the AAI (around 1 Hz), an online filter inducing temporal smoothing was applied to the AAI signal. The temporally smoothed AAI signal is defined as follows:

where W(τ) is a linear weighting function, i.e.,

where Δτ is the temporal support of the weighting function. Note that since , the fAAI signal represents a weighted average of the unfiltered AAI signal. As stated previously, the AAI signal is evaluated every 100 ms. Thus, in practice, only finitely many, discrete values of the AAI signal are available. Consequently, in the equation above, the AAI(.) function is substituted with a piecewise linear approximation that interpolates the discrete values.

Following recent guidelines, psychological variables that might influence training performance and could thus predict regulatory success were assessed using self-report measures [17,49]. These analyses will be reported separately. At the beginning of each session, the motivation to participate in the training was assessed using a seven-point Likert scale ranging from 1 (“very low”) to 7 (“very high”). In addition, we assessed the general self-efficacy, the health-related locus of control, and the current positive or negative affect of participants at the beginning of each session. This was done using the German versions of the General Self-Efficacy Scale (GSE) [50], the Multidimensional Health Locus of Control (MHLC) Scales [51], and the Positive and Negative Affect Schedule (PANAS) [52]. After each condition, i.e., twice per session, the perceived task demand and the effort exerted was assessed using the same seven-point Likert scale as for motivation and potential regulatory strategies were investigated by asking participants to “describe in a few words which strategy/strategies you have used to regulate the feedback signal.”

Before submitting data sets to the preprocessing pipeline, data were transformed to BIDS format [42] using the MATLAB-based (version: R2020b, Mathworks, Natick, Massachusetts, USA) toolbox FieldTrip (version: 20210212, [47]). Subsequent preprocessing of EEG data was conducted using the automatic preprocessing part of the DISCOVER-EEG pipeline [53–55]. In the Stage 1 manuscript of this Registered Report, we proposed a manual preprocessing pipeline. In consultation with the editorial board, and, importantly, before preprocessing any data sets, we switched from manual to automatic preprocessing, as this is the timelier method and enhances the objectivity, replicability, and speed of preprocessing. For each session, the preprocessing pipeline was applied to the concatenated neurofeedback data sets (ART_NF and ALT_NF or ART_sham and ALT_sham). Thus, 2 data sets were preprocessed for every participant. For every data set, intervals from −18 to 2 s with respect to laser stimuli were selected and concatenated for preprocessing. Preprocessing included line noise removal, bad channel detection and interpolation, re-referencing to the average reference, independent component analysis (ICA), and the automatic detection of bad segments. Prespecified control analyses (see below) assessing the impact of artefacts on the results of primary analyses confirmed that the amount of artefacts did not vary systematically between conditions (BF₁₀ = 0.41 and 0.13, respectively). Further analyses were performed using FieldTrip along with custom written code.

Primary analyses assessed the effects of our neurofeedback training in sensor space. Specifically, we examined distinct result patterns capturing attention, neurofeedback, and time effects (Fig 1) on alpha oscillations (hypothesis 1) as well as pain perception and underlying brain responses (hypothesis 2). In addition, we aimed to employ moderated multilevel mediation analysis to test whether changes in attention, alpha oscillations, and pain perception can be integrated into a single, mechanistic model (hypothesis 3).

Hypothesis 1: Alpha oscillations are up- and down-regulated. To examine changes in alpha oscillations, AAIs were extracted from the preprocessed neurofeedback data sets following the same procedure as during neurofeedback. For each trial, data from the last 12 s of the training period were extracted to limit the impact of the smoothing function used for AAI visualization (see section “Image visibility” in the Supporting information for details). Then, data were segmented into 1,000 ms epochs with 900 ms overlap and epochs with artefacts that had been marked during preprocessing were rejected (see section “Preprocessing” for details on applied artefact criteria). Remaining epochs were demeaned and power estimates from 8 to

Subsequently, asymmetry indices were analyzed using an adaptive analysis pipeline which allowed us to differentiate between 4 response patterns describing attention, neurofeedback, and time effects (see Table 1 and Fig D in S1 Text for a summary and visualization of the pipeline). As a basis for these analyses, single trial asymmetry indices of the first and second half of each condition (≙ 20 trials) were averaged separately for every participant resulting in 8 averages per participant (2 time periods × 4 conditions). To examine changes in alpha asymmetry over time (time effect; H1.1), single-subject first and second half averages were compared separately for ART_NF and ALT_NF conditions. If at least one comparison yielded a time effect (increase for ART_NF and a decrease for ALT_NF, respectively), second half averages were used as basis for all subsequent analyses. If no time effect was found, first and second half averages were averaged, and the resulting values were used for subsequent analyses. To examine whether verum neurofeedback led to larger increases/decreases in alpha asymmetry than sham neurofeedback (neurofeedback effect; H1.2), single-subject asymmetry averages from the verum conditions were compared to the corresponding sham conditions. If at least one comparison yielded a neurofeedback effect, data of the sham conditions would be used to evaluate whether ART and ALT conditions yielded opposite patterns of brain activity (attention effect; H1.3). Otherwise, data from corresponding verum and sham conditions was averaged, and the resulting single-subject ART and ALT average scores were compared.

Hypothesis 2: The perception of pain and underlying brain responses are up- and down-regulated. To examine changes in pain perception, pain ratings were extracted from the preprocessed neurofeedback data sets and were analyzed by repeating the analysis steps used for alpha asymmetry indices (including decisions on data selection, see Table 1 and Fig E in S1 Text for a summary and visualization). Since alpha oscillations and pain ratings are inversely related, the direction of pairwise comparisons was reversed, however. Subsequently, neurofeedback effects (H2.1) and attention effects (H2.2) on pain ratings were evaluated. Time effects were not analyzed due to confounding effects of habituation or sensitization.

Correspondingly, neurofeedback (H2.3) and attention effects (H2.4) were evaluated for brain responses which are typically observed after a noxious stimulus (see Table 1 and Fig F in S1 Text for a summary and visualization). These include a characteristic sequence of evoked potentials referred to as N1, N2, and P2 responses [56,57]. In addition, noxious stimuli suppress neuronal oscillations in the alpha (8 to 1,2]. Single-trial evoked and oscillatory brain responses to noxious stimuli were quantified using established procedures [46,58] which have been validated on a published data set ([59]; Fig F in S1 Text). To examine evoked brain responses, preprocessed data from the neurofeedback runs was bandpass filtered between 1 and 30 Hz (fourth-order Butterworth). Then, a baseline correction was applied using the fixation period preceding the neurofeedback training. Specifically, amplitudes from 2,000 to 2,500 ms with respect to the beginning of the fixation period were subtracted from the post stimulus data. Subsequently, evoked responses were quantified in a two-step procedure. First, individual peak latencies of evoked responses were determined based on averages across all trials of all neurofeedback conditions. To this end, local minima/maxima of the averaged waveform were determined at predefined channels (N1: C4, N2: Cz; P2: Cz) [46,58] and in predefined time-windows (N1: 120–200 ms; N2: 180–300 ms; P2: 250–500 ms) [46]. Second, single-trial amplitudes were obtained by averaging across a 30 ms window [58] centered at the previously defined peak latency. To quantify the N1 response, data were additionally re-referenced to Fz [60] before calculating the average waveform across all trials. To examine oscillatory brain responses, preprocessed data from the neurofeedback runs were filtered (fourth-order Butterworth 1 Hz high-pass filter, 41 to 51 Hz band-stop filter to dampen line noise). Next, single trial time-frequency estimates were obtained using a Hanning-tapered Fast Fourier Transformation and a sliding window approach. To obtain alpha and beta responses, a sliding window with a length of 500 ms and a step size of 20 ms was used. To obtain gamma responses, the window length was shortened to 250 ms, while the step size remained 20 ms. Finally, responses at alpha and beta frequencies were assessed by calculating the mean power across 8 to 12 Hz and 14 to 30 Hz, respectively, across a time window of 500 to 900 ms [46] and across the channels Cz, CPz, C2, C4, CP2, and CP4 [46]. Responses at gamma frequencies were assessed by calculating the mean power across 70 to 90 Hz [58], a time window of 150 to 350 ms [58] at Cz [46]. Brain activity in the theta frequency band was not analyzed as it mainly represents time-locked activity, which is captured by the laser evoked potential analyses [1,61].

Hypothesis 3: Changes in alpha oscillations mediate attention effects on pain perception. To link changes of alpha oscillations to changes in pain perception on a single-trial level, we employed a moderated multilevel mediation analysis based on data from all neurofeedback conditions. Mediation models examine whether the covariance between 2 variables (X and Y) can be explained by an intermediate variable termed mediator (M). Additional moderators make it possible to assess whether the strength of mediation effects varies across different conditions or groups [62]. Applied to the current study, such analyses allow us to assess whether the relationship between ART and ALT neurofeedback conditions (X) and single-trial pain ratings (Y) can be explained by the AAI (M) on a given trial and whether this effect is more pronounced for verum than for sham neurofeedback (moderator) (H3). To this end, the bi-variate relationships between X, M, and Y variables were quantified, and the mediation effect was determined based on the resulting path coefficients. Finally, path coefficients and the mediation effect can be compared between different levels of the moderator. Together, this procedure yields 2 advantages. First, mediation analyses quantify the relationship between single-trial brain activity (AAIs) and behavior (pain ratings) which is increasingly analyzed by neurofeedback studies [17]. Second, mediation analyses go beyond bi-variate brain–behavior analyses by integrating all variables of interest (i.e., attention, AAIs, and pain ratings) into a single model and thereby fosters mechanistic insights.

Control analyses assessed the impact of artefacts and unblinding on the results of primary analyses [17] by examining whether artefacts and blinding varied systematically between ART and ALT conditions. To this end, we calculated single-subject artefact (AI) and blinding indices (BI) for ART and ALT conditions and compared the resulting values between conditions. The A/BIs were defined as follows:

and

with trials_{AR/LT_NF/sham} representing the number of trials rejected in the respective condition and blinding_{AR/LT_NF/sham} representing the average blinding score in the respective condition.

All statistical analyses were conducted in the R programming environment (version: 4.1.1, [63]). Pair-wise comparisons were conducted using one-sided, Bayesian parametric paired samples t tests or nonparametric signed rank tests depending on the data distribution. For parametric t tests, the “BayesFactor” package was used (version: 0.9.12–4.2, [64], parameters: Cauchy prior distribution with a scale parameter r = √(2)/2 and truncation at zero). For nonparametric rank tests, the functions “signRankGibbsSampler” and “computeBayesFactorOneZero” were used ([65], parameters: Cauchy prior distribution with a scale parameter r = √(2)/2 and truncation at zero 1,000 iterations). Resulting Bayes factors (BF₁₀) are reported along with corresponding effect size estimates quantifying the median of the posterior Cohen’s δ distribution and its 95% credibility interval. Thereby, tests quantify evidence for both the presence and absence of effects [33].

Multilevel mediation analysis was conducted using the “bmlm” package (version: 1.3.4, [66,67]). Using Bayesian multilevel modeling and Markov chain Monte Carlo procedures, 5 path coefficients and corresponding confidence intervals were calculated for each participant (first-level coefficients) as well as on a group level (second-level coefficients). These coefficients represent the relationship between X and M (path a), the relationship between X and Y (path c), the relationship between M and Y controlled for X (path b), the relationship between X and Y controlled for M (path c’), and the mediation effect (path ab) [68]. To quantify moderation effects, we conducted 2 multilevel mediation analyses separately for verum and sham neurofeedback conditions. Subsequently, we planned to compare the resulting first-level coefficients between models separately for each path using the parametric or nonparametric tests described above under the condition that a mediation effect is found. Preceding the mediation analyses, X was recoded (ART, −1 and ALT, 1) and M (single-trial AAIs) was centered within-person [66].

First, we examined whether time up-/down-regulated alpha oscillations by comparing AAI values averaged within the first and second halves of each session. Specifically, we tested whether the averaged AAI would increase over time in the ALT_NF and decrease in ART_NF condition. The results provided moderate evidence that the AAI was higher in the second half of trials than in the first half of trials in the ART condition (BF₁₀ = 6.22; Fig 3B). In the ALT condition, the results provided moderate evidence against a difference in the AAI between the first and the second half of trials (BF₁₀ = 0.06; Fig 3B). Thus, time regulated alpha oscillations in one of the 2 conditions (ART), supporting hypothesis H1.1. Therefore, all subsequent analyses were based on averages across the second halves of trials.

We next investigated whether attention up-/down-regulated alpha oscillations. We found strong evidence that the AAI, which reflects the relative alpha power over the right somatosensory regions compared to left somatosensory regions, was higher in the ART than in the ALT condition (BF₁₀ = 51.90; Fig 3C). Thus, attention regulated alpha oscillations, supporting hypothesis H1.3.

We further assessed whether neurofeedback enhanced the attention-related up-/down-regulation of alpha oscillations. The results provided strong evidence that the AAI was lower in the ALT verum than the ALT sham condition (BF₁₀ = 25.71; Fig 3D). In the ART conditions, the results provided inconclusive evidence regarding an AAI difference between verum and sham conditions (BF₁₀ = 1.02). Thus, neurofeedback regulated alpha oscillations in one of the 2 conditions (ALT), supporting hypothesis H1.2.

Taken together, the results correspond most closely to predicted result pattern 1 (Fig 1), which incorporates effects of time, attention, and neurofeedback on alpha oscillations.

We first investigated whether attention regulated pain perception. The results provided moderate evidence against pain ratings being lower during ART (i.e., when focusing on the non-stimulated hand) than ALT conditions (BF₁₀ = 0.24; Fig 4B). Thus, although attention regulated alpha oscillations, it did not regulate pain perception, providing evidence against hypothesis H2.2.

We further assessed whether neurofeedback regulated pain perception. The results showed moderate evidence in the ART condition (BF₁₀ = 0.11) and strong evidence in the ALT condition (BF₁₀ = 0.09; Fig 4C) against a difference in pain ratings between verum and sham neurofeedback. Thus, although neurofeedback regulated alpha oscillations, it did not regulate pain perception, providing evidence against hypothesis H2.1.

We first investigated whether attention regulated brain responses to painful stimuli. The results provided moderate to strong evidence against differences in brain responses between ALT and ART conditions (see Fig 5C for results including BFs). Thus, although attention regulated alpha oscillations, it did not regulate brain responses to pain, providing evidence against hypothesis H2.4.

We further examined whether neurofeedback regulated brain responses to painful stimuli. The results showed moderate to strong evidence against differences in brain responses between neurofeedback and sham conditions (see Fig 5D for results including BFs). Thus, although neurofeedback regulated alpha oscillations, it did not regulate brain responses to pain, providing evidence against hypothesis H2.3.