2.1. Guideline 1: the rtfMRI signal is accurate and reliable

Necessary preconditions for any successful rtfMRI experiment are that the brain state of an individual is detectable and can be reliably and reproducibly converted into a feedback signal over the time-scale in question. Here we propose possible metrics that can be used to evaluate these prerequisites.

From a methodological perspective, the neurofeedback signal is generally derived from fMRI paradigms of two broad categories: general linear model (GLM)-based methods, and more recently, multivariate pattern analysis (MVPA) methods. For recent reviews of these methods, see Sulzer et al. (2013a) and LaConte (2011), respectively. GLM-based methods define an a priori ROI, either anatomically (using anatomical landmarks or atlas-based techniques) or with a functional localizer. The GLM is used to regress out nuisance parameters, and the resulting BOLD signal at each voxel in the ROI is combined into a neurofeedback signal using either averaging or a weighted average based on the standard deviation of the residual of the GLM in each voxel (Hinds et al., 2011). MVPA methods use supervised learning techniques, usually support vector machines, to determine the optimal set of weights (from either the whole brain or a restricted ROI) used to combine the BOLD signal across voxels into a single neurofeedback score.

In the context of rtfMRI, ‘real time’ is often used to describe both neurofeedback signals estimated from a single brain volume acquisition (Hinds et al., 2011) and across several brain volume acquisitions (Cox et al., 1995; Johnson et al., 2012; Yoo et al., 1999) to mitigate the contribution of noise in the BOLD signal. In order to estimate a reliable BOLD signal in a single measurement, it is important for models to include a moment-to-moment estimate of noise. One example of this approach is the incremental GLM method described in Hinds et al. (2011) and implemented in murfi2, software that is freely available on Github (http://github.com/gablab/murfi2).

Regardless of whether a GLM- or MVPA-based model is used to compute the neurofeedback signal or whether the signal is estimated in a single brain volume acquisition or across several acquisitions, the signal is a one-dimensional, usually linear combination of the BOLD signal across the brain. We can use the signal-to-noise ratio (SNR) to compute how well the neurofeedback signal conforms to the experimental design.

Let F represent the (stochastic) neurofeedback signal and X the (deterministic) experimental design vector. Under the standard GLM assumptions (Monti, 2011) SNR can be calculated by:

where corr(F,X) represents the Pearson correlation coefficient of vectors F and X. This quantity can be used to estimate how many repetition times (TRs) the neurofeedback signal would need to be averaged over to ensure it is accurate with a confidence level of (1 − α_F).

After theoretical modeling, we arrived at an equation that relates three key quantities: the confidence in the neurofeedback signal (1 − α_F), the signal to noise ratio (SNR_F), and the number of TRs used to compute the neurofeedback signal (n). The relationship between these 3 variables is:

where P() is the standard cumulative normal distribution and n is the length (in TRs) of both the task and fixation blocks.

Using this formula, we can determine the block length necessary to estimate an accurate neurofeedback signal with 95% confidence. We tested this formula using three rtfMRI feedback paradigms with regional ROIs with small (ventral striatum or VS), medium (fusiform face area and parahippocampal place area or FFA/PPA), and large (somatomotor cortex or SMC) expected SNR values. The results are illustrated in Fig. 1 and summarized in Table 1.

Open in a separate window

Fig. 1

The relationship between signal-to-noise ratio (SNR) and block length (n) required for a type I error rate (α) less than 0.05 for three ROIs.

Equally important to a detectable signal is a reproducible signal. Here the concordance correlation coefficient can be used to determine the reproducibility of the rtfMRI signal for a given subject from one run to the next, given the model employed. The concordance correlation coefficient is a simple metric that has been applied to fMRI to evaluate repeatability of various models (Lange et al., 1999). Let W₁ and W₂ represent the weights used to aggregate the feedback signal (derived from either a GLM or MVPA-based model) from runs 1 and 2, respectively. The concordance correlation coefficient between these weight vectors is then:

${\rho }_c=\frac{2\text{cov}(W_1,W_2^{})}{\text{var}\left(W_1\right)+\text{var}\left(W_2\right)+{({\overline{W}}_1-{\overline{W}}_2)}^2}$

where cov(.) represents the covariance of the weight vectors, var(.) represents the variance and ${\overline{W}}_1$ represents the mean.

From this work, we determined that the FFA/PPA and SMC ROIs were feasible as neurofeedback brain regions to target, but VS with low SNR was not feasible. It should be emphasized that detection of activation in FFA/PPA and SMC ROIs is feasible relative to the VS ROI; however, there are still substantial challenges in estimating neural activity from a single noisy BOLD volume (Hinds et al., 2011) that should be considered when deciding whether moment-to-moment neurofeedback from an ROI, set of ROIs, or components will produce the intended study outcome. At a minimum, it is recommended that researchers establish the quality of the neurofeedback signal in an independent dataset using either the SNR and concordance coefficient approach described above or some analogous method. If an experimenter finds that a signal does not meet these minimum standards, efforts should be made to optimize the experimental parameters and/or target ROI(s) before collecting data for a larger planned study. In cases where neurofeedback occurs from a region or network with low SNR, e.g., (Sulzer et al., 2013b), reporting SNR and methods used to optimize SNR, will help guide future researchers targeting these brain areas for neurofeedback.

2.2. Guideline 2: rtfMRI neurofeedback leads to learning

One use of rtfMRI is to provide feedback aimed at inducing learning that is difficult to achieve, or is less efficient, using other methods. This can be contrasted with alternative uses of rtfMRI, such as triggering task events or optimizing task parameters, where the goal is not necessarily to produce a lasting learning effect. To understand the learning induced by rtfMRI neurofeedback, three aspects of this process should be considered: (1) How is learning measured? (2) Which mechanisms are responsible? and (3) How meaningful or lasting are the effects? We will describe these three aspects followed by a case study.

Learning can be said to occur when experience influences behavior and/or alters brain structure or function. In the case of rtfMRI neurofeedback, the relevant experience can consist of elements of the task, feedback about regional brain activation, multivariate patterns of brain activity, or connectivity, and cognitive processes that arise due to the task, feedback, or attempts to control feedback with strategies. The consequences of this experience can be assessed using various dependent measures linked to learning in more standard experiments in cognitive psychology and neuroscience. In behavior, learning can be reflected in improved perception (Fahle, 2002), memory recall and recognition (Yonelinas, 2002), anticipation/prediction (Bubic et al., 2010), priming (Tulving and Schacter, 1990), and motor action (Stadler and Frensch, 1998). In the brain (particularly fMRI), learning can be reflected in enhanced (Schwartz et al., 2002) or attenuated activation within sensory systems (Grill-Spector et al., 2006; Turk-Browne et al., 2008), activation in learning and memory systems (Brewer et al., 1998; Poldrack et al., 2001; Wagner et al., 1998), changes to the multivariate representational space of brain regions (Folstein et al., 2013; Schapiro et al., 2013), changes to functional connectivity (Buchel et al., 1999), increased gray matter volume (Draganski et al., 2004), and alterations in white matter (Zatorre et al., 2012). All of these measures are potential targets for rtfMRI studies seeking to induce learning with neurofeedback.

These changes in behavior and the brain reflecting learning can arise from different mechanisms. One proposed mechanism is reward-based skill learning via cortical-basal ganglia brain networks (Birbaumer et al., 2013). Feedback from a brain region may produce a type of instrumental conditioning, whereby activation of that region becomes rewarding. An increase in the activity of the region, and possibly additional inputs from the reward system and regions involved in cognitive control, may induce local plasticity. This plasticity could be reflected in alterations of the selectivity and circuitry of neurons in that region (Sur and Rubenstein, 2005). There may also be larger-scale consequences. The region/representation used as the basis for feedback may become more involved in general, or more selective for an ongoing task. This could be analogous to establishing a compensatory mechanism, as occurs naturally after brain damage or in aging (Bedny et al., 2011; Heuninckx et al., 2008). However, the region(s) being “compensated” for (i.e., initially involved but not used for neurofeedback) remain intact, and could possibly be recruited less over time. Other regions that implement cognitive strategies for controlling feedback may become engaged in addition to the region being targeted for neurofeedback and the other brain areas recruited by this target region(s).

These learning effects differ in several ways that will impact the likelihood of obtaining an effect, where the effect will be observed in the brain, and whether the effect will be manifested in behavior. Several characteristics of learning will determine the feasibility of an rtfMRI study. For instance, learning occurs over a range of timescales, from immediately in the case of priming to over weeks in the case of perceptual learning. Relatedly, effects persist for different durations depending on the type of learning and brain system involved, from milliseconds for adaptation in the visual system (Grill-Spector et al., 2006) to years for episodic memories consolidated in the association cortex (McClelland et al., 1995). Finally, learning procedures vary in terms of whether the effects generalize to other contexts, from being hyperspecific to the training context (Jiang and Song, 2005) to more flexible (Turk-Browne and Scholl, 2009). Considering these parameters when designing a study will be important for its success.

These three aspects of learning – dependent measures, neural mechanisms, and timing/generalization properties – allow rtfMRI neurofeedback studies to be classified, and opportunities for new research to be identified. Take, for example, a study of perceptual learning induced by rtfMRI (Shibata et al., 2011). The dependent measures in this study were improved (1) accuracy in identifying a trained visual target orientation and (2) multivariate classification of this orientation from early visual cortex. The proposed mechanism is that rtfMRI feedback encouraged participants to generate orientation-specific neural activity patterns in this region, which resulted in local plasticity. The properties of learning in this task are that it took 5–10 sessions of training over a month, the effects lasted at least that long, and the benefit of training was specific to one orientation at a particular contrast. Various aspects of this classification could be investigated in future studies, such as what other brain systems are responsible for creating the activity patterns in the visual cortex, and whether other forms of perceptual learning (e.g., Xiao et al., 2008) could lead to more generalized benefits.

2.3. Guideline 3: the training protocol is optimized for rtfMRI-based neurofeedback and learning

A variety of approaches have been used for training subjects to control their brain patterns via rtfMRI neurofeedback. As it is yet unclear which approach yields optimal learning, more work is needed to provide evidence-based guidelines for clinical trial design. In most studies, the nature of the neurofeedback signal is made explicit to the subject, but there have also been paradigms where training occurs covertly (Bray et al., 2007; Shibata et al., 2011). Some studies present feedback while subjects are processing auditory, visual, or tactile stimuli (Bruhl et al., 2014; deCharms et al., 2005; Scheinost et al., 2013; Yoo et al., 2007) or while they are performing an assigned cognitive task (Chiew et al., 2012). Other rtfMRI studies employ an unconstrained paradigm in which the neurofeedback and the cues to increase or decrease brain activity are the only stimuli provided and the subjects are free to use a variety of cognitive strategies to control brain function and neurofeedback (Caria et al., 2007; Hampson et al., 2011; Rota et al., 2009). The optimal approach is likely dependent upon the application.

One open question is the importance of providing initial support (e.g., neurostimulation, pharmacotherapy, computerized cognitive training, and/or cognitive strategies) to subjects that will enable them to exert some initial level of control over the relevant brain activity patterns. Early reports suggested that learning was greatly facilitated by providing cognitive strategies to the subjects before they began neurofeedback (Caria et al., 2007; deCharms et al., 2005). Generally, rtfMRI neurofeedback studies have either provided all subjects with strategies or have not provided any subjects with strategies, making it impossible to determine the degree to which discussing strategies with subjects before they began neurofeedback helped the subjects to gain control over their patterns of brain activity. There have been no published studies to date that have used neurostimulation, pharmacological aids, or computerized cognitive training to enhance subjects' ability to utilize rtfMRI feedback, which may be an important avenue for future studies to explore.

It is not always necessary to provide initial strategies to subjects in order to achieve learning. Successful learning was achieved in subjects who were trained using instrumental conditioning (Bray et al., 2007) or were not provided with any initial strategies (Shibata et al., 2011). Further, subjects who have learned to control their brain activity patterns via rtfMRI neurofeedback have not always been consciously aware of the mental functions that were being molded by the training, even after completion of the neurofeedback (Shibata et al., 2011). These data imply that rtfMRI neurofeedback can induce subconscious learning. Although this is encouraging in terms of the potential utility of rtfMRI for training mental function, it also has ethical implications that must be carefully considered.

It is also important to consider the schedule of delivery of neurofeedback, whether it is continuous or intermittent, in order to optimize learning. Continuous neurofeedback or feedback delivered as soon as data is acquired, generally every TR (~1–2 s), has the advantage of delivering somewhat immediate (relative to the 2–6 s hemodynamic delay) feedback, which may be important to aid in efficient learning. For continuous rtfMRI neurofeedback the hemodynamic delay must be accounted for in the feedback display or explained to subjects. A disadvantage of continuous neurofeedback is that single measurements are often noisy, leading to the potential for inaccurate feedback if delivered on a TR-by-TR basis.

With continuous neurofeedback, subjects are asked to simultaneously attend to the task and the feedback signal, which increases cognitive load for the subject, and may disrupt optimal learning (van Merriënboer and Sweller, 2005). Other brain systems will be engaged by feedback that may be unrelated or even counterproductive to the task or training. For example, in a study by Greer et al. (2014), subjects were better able to use neurofeedback to increase nucleus accumbens activity by visualizing exciting events than to decrease activity by visualizing boring events. Subjects may have been less capable of decreasing nucleus accumbens activity with neurofeedback because the reward response to successful self-regulation of brain activity conflicted with the concurrent task of visualizing neutral or non-arousing events. In such cases, it could be advantageous to delay neurofeedback.

Additional brain regions appear to be recruited related to neurofeedback itself which have not yet been well-characterized. A study by Haller et al. (2013) reported increased functional connectivity during neurofeedback between an auditory target region and low level visual, insula, and working memory networks that was not found during a transfer phase with no feedback. Another study found activations in a broad frontoparietal and insula network as well as a broadly distributed negative network when subjects controlled neurofeedback as compared to viewing the feedback with no control (Papageorgiou et al., 2013). Control over feedback was further associated with improved whole-brain task signal-to-noise and increased pattern classification accuracy. Whether additional brain regions activated during neurofeedback are related to attention, control of feedback, or other factors such as learning and memory, has yet to be determined, including how these brain regions interact to produce neurofeedback-related brain changes. This work highlights the need for studies which investigate brain changes during continuous neurofeedback to consider the confounding effects of the neurofeedback (and learning) itself. Finally, there is evidence that continuous feedback may interfere with the consolidation of a learned response. Animal and human studies of operant learning have shown improved learning when subjects are provided with a period of delay after reward for “post-reinforcement synchronization” (Sherlin et al., 2011).

Intermittent neurofeedback or neurofeedback delivered after data acquisition over several TRs, generally between 8 and 60 s, allows for averaging of the feedback signal to improve SNR and accommodate the hemodynamic delay, and also minimizes cognitive load and other potential confounds by separating task strategy from the evaluation of feedback. Johnson and colleagues found that subjects were better able to manipulate activity in the premotor cortex when imagining movements with intermittent neurofeedback (at the end of each 20 s block) as compared to continuous neurofeedback (Johnson et al., 2012). However, the optimal delivery of neurofeedback likely depends on the specific application. While intermittent neurofeedback might be more effective by reducing cognitive load in studies in which individuals are provided with specific practice strategies, continuous neurofeedback might be useful to train individuals to fine-tune cognitive strategies related to specific patterns of brain activity.

Finally, it is important for researchers to consider human–computer interface design principles, especially as they relate to the display of the neurofeedback signal in order to aid in effective learning. This is a critical consideration as a poor human–computer interface design alone could lead to failed trials, or in some cases, adverse consequences such as increased frustration, confusion, and/or fatigue. No studies have been conducted to evaluate or optimize an rtfMRI neurofeedback interface design. A review of the human–computer interface design is beyond the scope of this paper (see Brown, 1998 for a general overview of this topic; Wickens et al., 2004); however, one area where a human–computer interface design may be helpful is choosing the optimal modality for delivering neurofeedback. Most rtfMRI neurofeedback paradigms, to date, have used visual feedback, but some subjects and populations may benefit from auditory, haptic, virtual reality/immersion, or some combination of these modalities for neurofeedback. In one application, an EEG-based neurofeedback signal was displayed on a participant's head using 3-dimensional cameras and a mirror to create a more realistic individualized display of brain function in real time (Mercier-Ganady et al., 2014). Future studies might query user experience (e.g., by questionnaires) to evaluate the functionality of the neurofeedback and the ease to which the user could learn to control the feedback signal, and compare this with a measure of accuracy with which the user could control the feedback, i.e., performance.

2.4. Guideline 4: there is an appropriate test of training success

It is also important to establish guidelines for how best to assess rtfMRI training success. To date, there have been two common approaches: subjects show improved (1) control of brain activity while receiving neurofeedback and/or (2) control of brain activity without neurofeedback (i.e., by comparing brain regulation without neurofeedback before and after training with feedback). A few studies have tested for transfer from training runs in which cognitive strategies were provided to control the feedback, to transfer runs in which cognitive strategies were used to control brain activity without any fMRI information. Young and colleagues reported that after several runs of neurofeedback training to increase activity in the left amygdala, subjects were able to significantly increase activity in the left amygdala using only cognitive strategies without neurofeedback (Young et al., 2014). In another recent study, Robineau et al. (2014) trained subjects to control interhemispheric visual cortex balance over 3 hour-long neurofeedback training sessions comprised of both feedback and transfer runs. They found that participants who learned to control the feedback signal were able to maintain control during transfer runs with no neurofeedback. In other studies, subjects had success self-regulating brain activity with the aid of neurofeedback, but they were not able to significantly self-regulate activity when no longer receiving neurofeedback (Berman et al., 2013; Caria et al., 2007; Greer et al., 2014; Hamilton et al., 2011). When creating an operational definition for training success, it will be important to consider: the expected timescale of training effects, expected pattern of change (e.g., linear or non-linear, monotonic or non-monotonic), and how best to account for individual differences. For example, training effects may be observed immediately or following some delay depending on the type and nature of learning impacting or impacted by training. Successful training may occur via gradual, incremental improvement characterized by a linear, monotonic function or trial-and-error testing characterized by a non-linear, non-monotonic function. It is also unclear whether the experimenter should fix the training interval or allow for adaptive training based on individual differences in optimal learning strategies and performance. Finally, the experimenter will need to design the study to adequately capture potential brain changes related to training over time, which could include reduced activity in the target ROI(s), change in the extent of activation within the ROI(s), and/or recruitment of different neural systems to support improved performance. The experimenter could consider including a resting state fMRI scan before and after training as one strategy for capturing complex brain changes over time (Hampson et al., 2011; Scheinost et al., 2013).

In order to determine whether rtfMRI neurofeedback training-induced changes are clinically significant, it will be important to introduce a metric of change such as the reliable change index (Jacobson and Truax, 1991) that accounts for measurement error and determines how functioning compares to a normative population. This is often used in clinical and neuropsychological assessments/studies where clinical significance is a question. Neuropsychological and clinical measures used for this type of an assessment typically have established psychometric properties (normative data, test–retest reliability, etc.), which make calculating reliable change indices possible. As the Human Connectome Project and other large neuroimaging databases are developed and data is shared, normative databases could be developed for fMRI data in order to calculate a reliable change index for neuroimaging data.

2.5. Guideline 5: rtfMRI neurofeedback leads to behavioral change

The behavioral effects of rtfMRI training may be manifest in improvement on the task used during rtfMRI neurofeedback training, improvement on related tasks or on the same task in other contexts, or improvement that generalizes to real-world outcomes (for review, see Ruiz et al., 2014).

In some studies, subjects have been trained to self-regulate brain activity by manipulating the neurofeedback signal and a behavioral response to some other, often concurrent stimulus. For example, deCharms and colleagues reported that training self-regulation of activity in the dorsal anterior cingulate cortex, a brain region implicated in pain perception and regulation, led to a corresponding change in the perception of pain caused by a noxious thermal stimulus as well as in spontaneous pain perception in patients with chronic pain (deCharms et al., 2005).

Another approach is to assess behavioral change before and after rtfMRI feedback training. For example, Zhang and colleagues trained subjects to increase activation in the left dorsolateral prefrontal cortex, a brain region involved in working memory, and reported improvements on digit span and letter memory tasks across training, indicating improved verbal working memory with rtfMRI neurofeedback training (Zhang et al., 2013). However, another study by Lawrence and colleagues found that although rtfMRI feedback could be used to train subjects to increase activity in the right anterior insula, training did not lead to changes in skin conductance response nor subjective valence ratings across pre- and post-training affective probes (Lawrence et al., 2013). Because so few evidence-based guidelines exist for rtfMRI studies, it is difficult to determine whether a lack of behavioral change is related to the specific brain–behavior relationship tested, or to methodology, such as only one rtfMRI training session in that study.

There has been limited work in non-clinical populations targeting emotional brain regions and function. Rota and colleagues have shown improved detection of emotional tone with rtfMRI training from the right inferior frontal gyrus in healthy participants (Rota et al., 2009). On the other hand, another study (Johnston et al., 2011) found that rtfMRI training from brain regions involved in positive emotions failed to improve mood ratings in healthy participants. Although methodology must be considered, the authors suggest that rtfMRI training of emotional control to enhance mood may be most effective in individuals with abnormal emotional control (as in their prior study in depression, Linden et al., 2012), and less effective in individuals who are capable of normal mood regulation (but see Allen et al., 2001, for a study demonstrating improved mood in healthy controls using EEG neurofeedback). A recent study of individuals with major depressive disorder (Young et al., 2014) found that rtfMRI neurofeedback training to increase activity in the amygdala during positive autobiographical memory recall led to improved self-reported mood post-scan compared to controls.

There is limited evidence of behavioral change from rtfMRI that has generalized to other tasks or real-world outcomes. Prior studies in clinical populations have shown decreased pain ratings in individuals with chronic pain (deCharms et al., 2005), decreased symptoms in individuals with tinnitus (Haller et al., 2010), decreased craving ratings and physiological response to smoking cues in nicotine-dependent individuals (Canterberry et al., 2013; Hanlon et al., 2013), decreased mood symptoms in people with depression (Linden et al., 2012), increased motor speed and clinical ratings of motor symptoms in individuals with Parkinson's disease (Subramanian et al., 2011), and decreased contamination anxiety in people with sub-clinical anxiety (Scheinost et al., 2013). However, with one exception (Scheinost et al., 2013), these studies measured behavior at the time of the rtfMRI study, did not test retention, and, with few exceptions (Scheinost et al., 2013; Subramanian et al., 2011) the dependent measure of behavioral change with rtfMRI was assessed with self-report measures that may reflect non-specific training effects.

Moving forward, it will be important for experimenters to show a causal link between the brain region(s) trained and the behavior targeted for modification in order to establish that specific brain–behavior changes account for any observed clinical changes and are not simply epiphenomena. It is a challenge to demonstrate cause-and-effect beyond correlation between rtfMRI neurofeedback training and behavioral change. Related to this, researchers should consider the specificity of the hypothesized relationship between the brain activation pattern(s) trained and behavioral changes. In the study by Shibata et al. (2011) described above, for example, causality was established between rtfMRI neurofeedback training from the early visual cortex corresponding to a specific visual orientation and improved accuracy in identifying the target orientation as compared to untrained orientations. In such studies the specific brain activation pattern related to behavior must first be determined, for example using multivariate pattern analysis as in Shibata et al., (2011) and as described above (LaConte, 2011). In another recent study, Zhao et al. (2013) used dynamic causal modeling with rtfMRI to measure causal interactions between the dorsal premotor cortex target region and other motor regions during a motor imagery task with neurofeedback. The experimental group showed increased interactions from the target brain region to the other regions across four training sessions as compared to a sham control group. The experimental group also showed improved performance on the motor imagery task. rtfMRI methods need to be further refined to establish causality not only between neurofeedback training and changes in behavior, but also between changes in the specific neurofeedback signal and other brain changes.

2.6. Guideline 6: an appropriate rtfMRI neurofeedback-based clinical trial design is in place

Despite important early work suggesting that it is possible to use rtfMRI as a non-invasive brain-based clinical tool, to our knowledge there has been only one RCT, recently completed, that has investigated the efficacy of rtfMRI neurofeedback to effect meaningful clinical change (CH, personal communication). Several methodological considerations for the design of rtfMRI studies and clinical trials remain open questions that likely depend on the specific application.

To demonstrate behavioral change that is directly related to rtfMRI feedback training and establish causality, studies must implement important control conditions. Most studies that have included a control condition have either used false feedback or no feedback. False feedback can involve providing subjects with arbitrary feedback not related to brain function, actual neurofeedback from a brain region or network theoretically unrelated to the experimental variables of interest either within-group (e.g., Garrison et al., 2013a) or between groups (e.g., Lawrence et al., 2013; Young et al., 2014), or yoked neurofeedback from a matched subject (e.g., Hampson et al., 2012). In a recent study, false feedback based on a recording of EEG neurofeedback was found to engage a broad network of frontal, parietal and cingulate regions involved in cognitive control (Ninaus et al., 2013). In another study, fixed randomized feedback not based on an actual fMRI signal was compared to no feedback as a control for training self-regulation of activity in the premotor cortex (Johnson et al., 2012). False feedback again produced a widespread pattern of activation involving frontal, temporal, and parietal regions, which was distinct from the more localized activation associated with actual neurofeedback. In addition, subjects reported more frustration with the task in the false feedback group as compared to the no feedback or actual neurofeedback groups. Based on these findings, the authors reasoned that the negative impact of false feedback runs made it a less suitable control group than no feedback. However, the drawbacks to providing no feedback are first that it is unlikely to be as engaging as a feedback task, and second, that it does not control for the perception of success that subjects experience when they do well in controlling their brain patterns. These differences between a neurofeedback group and a no feedback control group can lead to false positives related to unmatched motivation and placebo effects. Another option is to deliver control feedback using more cost-effective methods such as autonomic biofeedback (deCharms et al., 2005).

Other important design considerations for rtfMRI clinical trials include the optimal number of rtfMRI sessions, number of neurofeedback runs per session, appropriate timing between sessions if multiple sessions are used, and the combination of rtfMRI training with behavioral or other interventions. There is little data available to guide optimization of these parameters for clinical trials. A recent study addressed a number of these issues by tracking change across three rtfMRI sessions in which nicotine-dependent individuals were trained to reduce activation in the anterior cingulate cortex and reduce smoking cue-related craving (Canterberry et al., 2013). Of 15 enrolled smokers, 60% completed three 1-hour rtfMRI sessions, 1–2 weeks apart. Within each rtfMRI session, subjects completed three 10-minute feedback runs. Reduced anterior cingulate cortex activity and reduced self-reported craving were evident at the first rtfMRI session and consistent across sessions and runs. The reduction in cue-induced craving with rtfMRI neurofeedback was significant at the third session, indicating that at least two feedback sessions were necessary to see any effect of neurofeedback, and more than two sessions may be needed to observe clinical improvement.

Finally, it will be critical to compare the effects of rtfMRI-based neurofeedback to existing therapies or biofeedback using more cost-effective neuroimaging tools such as EEG in order to demonstrate the value added by rtfMRI-based neurofeedback above other treatment options.

The ideas for this paper were developed at a workshop in July 2013 supported by the Radcliffe Institute for Advanced Study. Support was also provided in part by the National Institutes of Health: K23DA032612 (LES); K24DA030443 (AEE); R21DA030523 (AEE, JDG); R21/33DA026104 (HCB, AK); K01 DA034093 (JMG), K24DA029262 (SM), R01 AT007922-01 (JAB), R03 DA029163-01A1 (JAB), K12 DA00167 (JAB), P50 DA09241 (JAB), R01 AA02152901A1 (SML), R01 MH095789 (MH), 5R21DA026085 (MG), the Norman E. Zinberg Fellowship in Addiction Psychiatry at Harvard Medical School (LES, JMG), philanthropic donations from the Poitras Center for Affective Disorders Research (SWG), the Charles A. King Trust (LES), the McGovern Institute Neurotechnology Program (SG, JDG, AEE), and private funds to the Massachusetts General Hospital Department of Psychiatry (AEE). Research presented in this paper was also carried out in part at the Athinoula A. Martinos Center for Biomedical Imaging at the Massachusetts General Hospital, using resources provided by the Center for Functional Neuroimaging Technologies, P41RR14075, a P41 Regional Resource supported by the Biomedical Technology Program of the NIH National Center for Research Resources (NCRR).