Can we predict real‐time fMRI neurofeedback learning success from pretraining brain activity?

Abstract Neurofeedback training has been shown to influence behavior in healthy participants as well as to alleviate clinical symptoms in neurological, psychosomatic, and psychiatric patient populations. However, many real‐time fMRI neurofeedback studies report large inter‐individual differences in learning success. The factors that cause this vast variability between participants remain unknown and their identification could enhance treatment success. Thus, here we employed a meta‐analytic approach including data from 24 different neurofeedback studies with a total of 401 participants, including 140 patients, to determine whether levels of activity in target brain regions during pretraining functional localizer or no‐feedback runs (i.e., self‐regulation in the absence of neurofeedback) could predict neurofeedback learning success. We observed a slightly positive correlation between pretraining activity levels during a functional localizer run and neurofeedback learning success, but we were not able to identify common brain‐based success predictors across our diverse cohort of studies. Therefore, advances need to be made in finding robust models and measures of general neurofeedback learning, and in increasing the current study database to allow for investigating further factors that might influence neurofeedback learning.

However, not every individual can benefit from neurofeedback training and neurofeedback learning success differs substantially between individuals. In fact, many studies report participants who were unable to gain control over their own brain activity, even after multiple training sessions. In these studies, an average of about 38% of all participants failed to modulate their own brain activity and were not able to reach predefined goals after neurofeedback training (Bray, Shimojo, & O'Doherty, 2007;Chiew, LaConte, & Graham, 2012;deCharms et al., 2005;Johnson et al., 2012;Ramot, Grossman, Friedman, & Malach, 2016;Robineau et al., 2014;Scharnowski et al., 2012;Yoo et al., 2008). This failure to modulate brain activity, also referred to as the "neurofeedback inefficacy problem" (Alkoby, Abu-Rmileh, Shriki, & Todder, 2017), leads to a reduction in overall efficiency of neurofeedback training and hampers translation to clinical interventions. To date, the factors that cause neurofeedback inefficacy as well as the large inter-individual variability in neurofeedback learning success in the field of real-time fMRI neurofeedback remain unknown.

| Received data on pretraining activity and neurofeedback learning success
We asked the authors to provide one value determining neurofeedback success for each neurofeedback training run, and one value determining pretraining brain activity levels within the ROI that was trained during neurofeedback. In particular, we asked for individual data for each participant of an experimental neurofeedback training group, excluding control groups such as receiving sham feedback or modulating brain regions of no interest. Most contributions consisted of data that were already fully analyzed and published. For most studies, we also received fully processed beta values for average pretraining activity levels within the trained ROI. In some cases, we extracted these values using target ROI masks and contrast images of the corresponding pretraining run [3,6,9,10,18], or from raw data [7, 14].

| Data analysis of raw data
For the study that shared the raw data, we analyzed the data using a standard preprocessing procedure in native space (slice time correction, motion correction, coregistration, spatial smoothingwith a Gaussian kernel of 6 mm full width at half maximum, no normalization) using Due to small sample sizes, further subdivisions of the data in (4) and (5)  For each of these five groups as well as the entire sample (all data from all studies), we calculated overall meta-correlations using a weighted (weights based on the number of participants included in the study) random-effects model. All statistical meta-analyses were performed using the meta package in R using the metacor function (www. cran.r-project.org/web/packages/metacor). Studies that included both patients and healthy subjects, and studies that investigated both connectivity-and activity-based neurofeedback were split into several corresponding sub-groups accordingly. One study that trained a different ROI for each participant [21] was not considered in the ROI-based group split. Further, some of the studies included in the no-feedback group or the ROI-engaging paradigm group included a functional localizer scan in their experimental design but, due to data dropouts, the corresponding no-feedback or ROI-engaging paradigm runs were used to extract activity levels. In addition, we performed several analyses to quantify heterogeneity of effect sizes using the Meta-Essentials tool (Suurmond & Hak, 2017).

| Activity-based neurofeedback with healthy subjects
For activity-based neurofeedback with healthy subjects, we found no significant relationship between pretraining activity levels and F I G U R E 2 Averaged weighted Spearman correlations between pretraining activity levels and neurofeedback learning success as measured by the slope of the learning curve. Circle sizes represent the corresponding study's sample sizes. Further, the coloring scheme reflects the corresponding grouping of the subjects (healthy subjects/patients) and the studies (type of feedback, trained target region(s) and type of pretraining activity levels). Overall, no correspondence between pretraining activity levels and neurofeedback learning success was found. Abbreviations: amy, amygdala; DMN, default mode network; n.a., not applicable; no-fb: no feedback; loc, localizer; ROI-eng, ROI-engaging For activity-based neurofeedback studies with healthy subjects, we also found no significant relationship between pretraining activity levels and success in the first neurofeedback run (r = −0.06, p = .49), with 6 of 12 studies even showing a negative correlation. Heterogeneity measures again showed low heterogeneity of effect sizes (Q = 12.45, Q-df < 0, p Q = 0.33, I 2 = 11.68%, T 2 = 0.01, T = 0.10).
F I G U R E 3 Averaged weighted Spearman correlations between pretraining activity levels and neurofeedback learning success as measured by the difference between neurofeedback success in the last and the first neurofeedback run. Circle sizes represent the corresponding study's sample sizes. Further, the coloring scheme reflects the corresponding grouping of the subjects (healthy subjects/ patients) and the studies (type of feedback, trained target region(s) and type of pretraining activity levels). Overall, no correspondence between pretraining activity levels and neurofeedback learning success was found, except for when only investigating pretraining activity levels during a functional localizer run. Abbreviations: amy, amygdala; DMN, default mode network; n.a., not applicable; no-fb, no feedback; loc, localizer; ROI-eng, ROI-engaging F I G U R E 4 Averaged weighted Spearman correlations between pretraining activity levels and neurofeedback learning success during the first neurofeedback run. Circle sizes represent the corresponding study's sample sizes. Further, the coloring scheme reflects the corresponding grouping of the subjects (healthy subjects/patients) and the studies (type of feedback, trained target region(s) and type of pretraining activity levels). Overall, no correspondence between pretraining activity levels and neurofeedback success in the very first neurofeedback run was found. Abbreviations: amy, amygdala; DMN, default mode network; n.a., not applicable; no-fb, no feedback; loc, localizer; ROI-eng, ROI-engaging 3.3 | Connectivity-based neurofeedback with healthy subjects

| Activity-based neurofeedback with patients
For activity-based neurofeedback studies across different patient populations, we did not find a significant correlation between pretraining activity levels and neurofeedback learning success, for neither neurofeedback learning success measures (slope of the learning curve: r = −0.13, p = .20; last vs. first rundifference: r = −0.14, p = 0.19).
Here, 6 of 8, and 7 out of 8 studies showed a slightly negative relationship, respectively. Heterogeneity of effects sizes was very low   Table S1), we did not find significant effects for any of the assessed functional domains, that is, amygdala (emotion processing), DMN/PFC (mind wandering and higher cognitive functioning), motor functioning, reward processing, and other sensory domains. For neurofeedback success measured by the difference between success in the first and the last neurofeedback run (see Table S2), we found a negative correlation for studies that focused on DMN/PFC regulation (r = −0.13, p < .001). We did not find significant effects for any functional domain clusters when investigating the correlation between pretraining activity levels and neurofeedback success during the first neurofeedback run (see Table S3).

| Type of pretraining run
Pretraining activity levels were either based on a no-feedback run, a functional localizer run, or on another task engaging the ROI that was not used for localizing the ROI, for example, a finger tapping task when neurofeedback training was targeting the motor cortex [13]. Overall, studies with a functional localizer run showed a significant positive correlation between the localizer activity levels and neurofeedback learning success as measured by the difference between neurofeedback learning success in the last and the first neurofeedback run (r = 0.12, p = .003). However, this correlation was not significant when success was measured by the slope of the learning curve (r = 0.09, p = .20). For activity levels during other pretraining runs we did not observe a significant correlation with learning success. Further, none of the three types of pretraining run groups showed significant correlations between pretraining activity and the very first neurofeedback run (see Tables S4-S6 for exact values).

| DISCUSSION
Here, we performed a meta-analysis with 24 different fMRI-based neurofeedback studies to investigate whether pretraining activity levels can be used to predict neurofeedback learning success. In our data set of 401 subjects undergoing neurofeedback training, we did not find an overall significant relationship between these two measures, that is, ROI activity prior to neurofeedback training and neurofeedback learning success were not significantly correlated.
One of the reasons for not having found an overall relationship between pretraining activity and learning success might be that the studies included in this meta-analysis are quite diverse in terms of, for

| Differences between healthy subjects and patients
Neither healthy subjects nor patients showed a significant correlation between pretraining activity levels and neurofeedback learning success.
Interestingly, the majority of patient studies showed a negative correlation between neurofeedback learning success and pretraining activity levels, while we observed more positive correlations for studies with healthy subjects. This might be explained by symptom severity being associated with increased ROI activity, which again can influence a patient's neurofeedback learning performance. For example, patients suffering from substance use disorder who show highly increased craving-induced brain activity levels might be less successful in downregulating craving-related brain signals than addiction patients who only show mildly increased craving-related brain activity. Increased brain activity levels in higher order brain areas might also be an indicator for decreased cognitive capacitiesas the performed task constitutes a particular challenge to the patients, they might experience exhaustion during the following neurofeedback training runs. Further, aspects like differences in adaptation, motivation, deficits in sustained attention etc.
that are often reported in specific patient populations, might also drive neurofeedback training success differences.

| Activity-versus connectivity-based neurofeedback
Neither activity-nor connectivity-based neurofeedback studies showed a significant correlation between pretraining activity levels regions within the brain. For instance, Bassett and colleagues suggest that highly connected brain regions such as areas within the DMN, might be easier to train than less-connected brain areas (Bassett & Khambhati, 2017). This is also in line with recent suggestions that connectivity-based measures might be more promising for predicting complex higher order cognitive processes than measures based on single brain regions (see Horien, Greene, Constable, & Scheinost, 2020 for a review on this topic). Indeed, several activity-based neurofeedback studies report concomitant changes in brain connectivity (Lee et al., 2011;Rota, Handjaras, Sitaram, Birbaumer, & Dogil, 2011;Scharnowski et al., 2014;Scheinost et al., 2013;Zotev et al., 2011;Zweerings et al., 2019).
Thus, future analyses should consider connectivity measures as predictors not just for connectivity-based neurofeedback studies, but also for activity-based studies.

| Type of pretraining run
Interestingly, when grouping together studies based on the paradigm of the run during which pretraining activity levels were collected, we observed a significant positive correlation between pretraining activity  (Haller et al., 2013), selfregulation performance (Robineau et al., 2014), and signal-to-noise ratio (Papageorgiou, Lisinski, McHenry, White, & LaConte, 2013). This also indicates that the feedback has a stronger effect on neurofeedback training success than the actual task the participant is performing in the scanner. Indeed, recent implicit neurofeedback studies show that neurofeedback learning is possible even when participants are not informed what the neurofeedback signal represents and are not provided with mental strategy instructions that are related to the function of the target ROI (Cortese, Amano, Koizumi, Kawato, & Lau, 2016;Koizumi et al., 2017;Shibata et al., 2011;Taschereau-Dumouchel et al., 2018). These findings show that neurofeedback runs are special in that they constitute their own specific experimental condition that is distinct from seemingly-related conditions such as transfer runs without neurofeedback. Thus, they should be analyzed separately and, for  (Rosenstiel & Keefe, 1983) and state anxiety scores (Spielberger, 2010) predict success in learning to regulate the ACC  and emotion networks , respectively, another study did not find correlations between pretraining spatial orientation (Stumpf & Fay, 1983), creative imagination (Barber & Wilson, 1978), or mood scores (Zerssen, 1976) scores and success in learning to regulate pre-motor and para hippocampal ROIs . A recent systematic review on psychological factors that might influence neurofeedback learning success in EEG and fMRI studies argues that factors such as attention and motivation might play an important role in successful neurofeedback training runs (Cohen & Staunton, 2019). However, although these are likely candidates for affecting neurofeedback learning, a concrete empirical effect of these factors has so far only been reported in one fMRI-based neurofeedback study (Chiew et al., 2012), showing a clear necessity for more empirical investigations on these factors.
In EEG neurofeedback, several factors have been observed to be correlated with neurofeedback learning success (Alkoby et al., 2017), but they were only reported in single EEG studies have not yet been tested in fMRI-based neurofeedback studies. For instance, factors that seemed to have a positive influence on EEG-based neurofeedback learning success were regular spiritual practice (Kober et al., 2017) or a relaxing attitude towards one's ability to control technological devices (Witte, Kober, Ninaus, Neuper, & Wood, 2013). Also, other brain-based measures that are, for example, focused on areas more generally involved in self-regulation (Emmert et al., 2016) or on connectivity rather than activity levels (Horien et al., 2020)might be suitable candidates that should be explored in future studies. The latter might be particularly relevant for connectivity-based neurofeedback studies, but we were not able to test this due to lack of suitable data.
Further possible candidates for predicting neurofeedback success might be factors that have already been identified to be predictive of cognitive and behavioral training success in non-neurofeedback studies, for example, activity in areas related to stimulus encoding and motor control has been found to be predictive of motor learning (Herholz, Coffey, Pantev, & Zatorre, 2016), and activity in the motor network has been found to predict training-related changes in working memory (Simmonite & Polk, 2019). Finally, very recent work by Skouras et al. indicates that neurofeedback learning performance can be influenced by biological factors such as genetic and anatomical predispositions , thus demonstrating the complexity of the underlying processes and the need for using multimodal data sets.
Hence, currently, no robust predictors for neurofeedback learning success have been identified, and, even if predictions can be made, they are likely study-specific (i.e., questionnaires that are specific to the trained ROI) and might not generalize across studies. Besides empirical studies, future studies using secondary mega-analyses might be a promising tool to identify factors that influence neurofeedback learning.

| CONCLUSION
Here, we aimed at finding general pretraining predictors for neurofeedback training success. We observed a slightly positive correlation between pretraining activity levels during a functional localizer run and neurofeedback learning success, but we were not able to identify common brain-based success predictors across our diverse cohort of studies. In order to achieve the goal of finding predictors for neurofeedback learning success advances need to be made: in developing (a) models for neurofeedback learning, (b) establishing robust measures for neurofeedback learning, and (c) in increasing the database including acquired candidate measures across numerous studies.
The reward of such a joint effort would be increased efficacy and costeffectiveness of this promising scientific and therapeutic method.