The ‘Go’s and the ‘No-Go’s of response-inhibition training to food: lessons learned from trials

High food-reward sensitivity and low inhibitory control are modifiable targets for overeating interventions. Our review of 16 food-related response-inhibition training (RIT) trials identified key elements linked to effectiveness, including recruiting from at-risk populations (i.e. those with overweight or heightened snacking behaviour), and designing intervention tasks to support bottom-up, associative (food-inhibition) learning. The optimal comparison condition depends on the research question, but the most consistent training effects have been seen relative to generalised (non-food) RIT. Trial outcomes should prioritise objective and validated measures (e.g. weight loss and explicit food devaluation). Future trials should consider unanswered questions such as training schedules and timing, and whether training people to ‘go’ to healthy foods can increase their appeal. It includes 1) a limited number of items/categories of HFSS food that are consistently and repeatedly paired with No-go cues. 2) A distinct category of healthy foods is always paired with Go cues and 3) a third distinct category of fillers (e.g. clothing items) is equally paired with Go and No-go cues. The fillers provide a baseline for measuring associative responses to food items, keep the task challenging and help to make the foods relatively more predictive.


Introduction
Poor diet is one of the leading causes of premature mortality [1]. Many people know they should eat more fruits, vegetables and wholegrains and fewer foods high in fat, sugar and salt (HFSS), but struggle to translate their intentions into behaviour. This is partly because HFSS foods are highly rewarding, and rapidly attract attention and elicit motor excitation and approach responses, resulting in cravings and impulses that can be hard to control [2]. Those at particular risk of overeating have strong automatic reward responses to HFSS food and ineffective response inhibition [3]. Recent interventions for overeating target these automatic responses to food, for example, by training people to attend away from HFSS foods or by repeatedly avoiding or inhibiting their motor responses to them [4].
In this narrative review, we focus on identifying the key features of 16 trials of food-specific response-inhibition training (RIT) [5• -19], which are associated with the most positive outcomes. In doing so, we identify the most suitable participants, intervention components, comparison condition, and outcomes (PICO) from realworld trials to guide the design of future food RIT interventions and evaluations.

Background summary
Food RIT adapts Go/No-go (GNG) or stop-signal tasks (SSTs) to require participants to respond (Go) to neutral items or healthy foods, and to inhibit responding (Nogo) to HFSS foods [20]. Specific stimuli (e.g. images of sweet snacks) are consistently paired with a stop or no-go signal, such as an auditory tone or a coloured frame around the image. This signal is presented simultaneously with HFSS food in GNG tasks, making motor inhibition easy, but is presented after a delay in SSTs, making inhibition more difficult (see [21] for differences between GNG and SST). Over the past decade, approximately 36 studies have examined food RIT, and several meta-analyses report robust medium effects (d = 0.4-0.5) on reducing food intake and choice [4,[22][23][24][25] (for relevant p-curve analyses see [26,27]). Stronger effects result from tasks that train more associative 'bottom-up' forms of inhibition, such as the GNG, relative to tasks that place greater demands on 'top-down' forms of inhibition, such as the SST [23,28].
The mechanisms underlying food-RIT effects on intake and choice are unclear, but evidence indicates a role for learnt (automatic) stimulus-inhibition associations [29,30] and reduced liking (devaluation) of inhibited food stimuli, another consistently observed effect of food RIT (e.g. [31,32•]), which is theorised to resolve conflict between 'go' and 'inhibit' response tendencies [33]. In contrast, there is scant evidence for training-related improvements in top-down inhibitory control as measured on separate GNG or SSTs, which generally show negative results [8,16,17]. Neuroimaging data support the devaluation and automatic motor-inhibition (reduced motor excitation) accounts, rather than improvements in top-down inhibitory control; studies show training-related reductions in activation in reward, attention and motor-related brain regions, but no increases in activation in prefrontal brain circuits associated with inhibitory control [9•,19], see also [34].
Whilst the laboratory evidence that food RIT modifies eating behaviour is robust [4,[22][23][24], it is unclear how well intervention effects translate to the real world and over the longer term. This review considers 16 trials examining the effects of multiple sessions of food RIT (see Table 1). Trials are identified based on the 'trial' (vs. lab study) inclusion criteria recommended by Wiers et al [35]. 1 The majority (12 out of 16) of these trials reported positive effects on at least one outcome related to eating behaviour (weight, food intake or liking, see Table 1). We have identified common features of these trials, and using the PICO framework, we provide a current guide to developing an effective food-RIT intervention and appropriate evaluation. These include the need to recruit at-risk populations, to employ tasks that train associative inhibition effectively (and show evidence of this) and to include both validated and objective outcome measures, for example, stimulus devaluation and weight loss.

Population
Both common sense and evidence suggest that trial participants should be those who like/are frequent consumers of the trained no-go foods (i.e. those for whom training-induced devaluation and inhibition could have the greatest impacts on energy intake and weight loss). Indeed, food RIT facilitates weight loss among participants with a higher body mass index (BMI) [14,18] or with greater baseline food intake or liking [11••,14]. Null effects on these outcomes have been observed in healthy-weight, unselected (e.g. low snacking) samples [5•], see also [36•] for similar findings using combined food GNG/cue-approach training. More proximal outcomes (such as devaluation) can be impacted in healthy populations [5•,36•], however, evidence indicates that food RIT brings the greatest behaviour change and weight-loss benefits to 'at risk' or overweight populations.

Intervention
RIT task parameters should be carefully designed to support bottom-up associative (HFSS food-inhibition) learning. This includes facilitating high accuracy on HFSS food-No-go trials, which is linked to RIT effects [23] (Box 1, Figure 1). Inhibition accuracy tends to be higher (see Table 1) in GNG tasks, likely because the task is easier than the SST. However, this needs to be complemented with ensuring the task remains challenging and engaging. RIT tasks should include at least 50% Go trials (to place some demands on inhibition, [31]) and include filler items that have 50% Go and 50% No-go response mappings [14], which also enables a direct measurement of stimulus-response (S-R) learning during training (e.g. greater No-go accuracy and faster Go reaction times to consistently paired foods vs. inconsistently paired fillers indicate associative learning). It is notable that all six trials showing evidence of such S-R learning in their training tasks reported significant effects on at least one of weight loss, food intake or devaluation [5•,7,13,14,16,19]. The remaining trials either did not measure or report RIT task performance data, but it is possible to estimate task potential for training related associative-inhibition at the HFSS food stimulus or category level by examining task parameters (see Box 1). We suggest that all of the trials reporting positive findings are likely to have trained associative inhibition to HFSS foods (Table 1), with the exception of one ([6] study 1), whose BMI findings were based on self-report and failed to replicate using objective measures ( [6] study 2). In contrast, it is doubtful whether the four food-RIT trials reporting null effects effectively trained associative inhibition to HFSS foods, for example, because they used a SST or included diverse Go and No-go foods, which may compromise category-level learning and generalisation ([6] study 2 [8,10,17]; see Table 1. See also [40] for a lengthier discussion of this point).
Interventions are only effective if the target audience engages with them. Lab studies have shown that the amount of attention paid to stimuli during RIT can determine the strength of effects on subsequent outcomes [41,42] and testing in busy/distracting environments can sometimes lead to null effects (e.g. [43]). Manipulating the GNG cues may help to boost attention to food stimuli, for example, by having a short delay between the presentation of the food and No-go cue, or by using the foods themselves as a stop cue (Box 1). Gamification to increase short-and long-term engagement is a potential solution (see [36•] for an excellent example of a trial testing food GNG with state-of-the-art game design principles). However, gamification should be executed carefully as it can weaken or obliterate training effects, for example, if complex visuals and mechanics interfere with attention to food stimuli and S-R learning [11••,17]. Gamification may be more Table 1 Summary table of key elements in trials of food-specific RIT describing the PICO. Trials with an asterisk showed positive findings for at least one eating-related outcome (weight loss, food intake or food valuation). The Intervention column includes (in italics) evidence for, or an estimate of, associative S-R learning. FFQ = food frequency questionnaire, VAS = Visual Analogue Scale. L/R = left/right response, EEG = electroencephalography, fMRI = functional magnetic resonance imaging, EMA = ecological momentary assessment, EDE-Q = Eating Disorders Examination questionnaire. beneficial for some populations such as children [43] and men, but not women [12]. We recommend implementing simple gamification strategies that enhance user engagement (such as point scoring) without compromising GNG training [44]. In sum, we recommend (i) designing tasks that challenge inhibition while allowing high accuracy, (ii) ensuring Go and No-go food categories are clearly defined, (iii) including measures of S-R learning in trial reports (e.g. by comparing accuracy and response times to foods vs. fillers), as without these data, null effects of RIT interventions are difficult to interpret and (iv) implementing simple gamification strategies to enhance engagement.

Comparison
The comparison conditions used in food-RIT trials have included active control training tasks, and no-training 'treatment as usual' conditions, sometimes involving psychoeducation [10,13,17].  45]). Most trials therefore opt for a non-food control training (e.g. [18]) and these have generally shown positive effects (with the exception of [8]) but participant blinding may not be optimal [5•].
One might expect the strongest effects in trials using a no-training control condition but this has not been the case so far, with only one of three such trials showing positive effects [13], although the other two trials [10,17] did not use optimal interventions, making interpretation difficult. We recommend that future trials continue to Figure 1 for an example of recommended methods.

Aim
Rationale and evidence Recommended methods (see figure 1 for example) Other methods

Establish S-R inhibition learning
Create direct associations between HFSS foods and the inhibition of a motor response.
(Differs from indirect associations between HFSS foods and a stop signal, which may not be consistently followed by response inhibition, [30]) Specific HFSS foods are repeatedly associated with the successful inhibition of a response. For example, GNG task where HFSS foods are 100% associated with a No-go cue (Figure 1, 1) and inhibition is easy resulting in high rates of successful inhibition (e.g. above 95% [13]) HFSS foods are inconsistently associated with response inhibition. For example, SSTs where HFSS foods are equally presented on Go and stop trials and the stop-signal delay is calibrated to maintain stopping accuracy at ∼50% (i.e. participants would inhibit responding to HFSS foods on ∼25% of HFSS trials) [6,10] 2. Establish category-response inhibition learning Encourage associative inhibition to generalise from the item to the category level (i.e. to HFSS foods other than the precise items presented during training) Several exemplars from a clear and meaningful category of HFSS food [47] or a small number of categories [13] are consistently paired with response inhibition (Figure 1, 1). A clearly delineated category of healthy food such as fruits and vegetables, for example, [14] or non-food items, for example, [18] is paired with Go responses (Figure 1, 2). Task may also include a third clearly delineated category of filler items such as clothing (Figure 1, 3 [14]).
No clear HFSS-food category is associated with response inhibition. For example, a wide range of healthy and unhealthy foods are paired with Go and No-go trials respectively and/or food items are not repeated between training sessions (e.g. [8]). This could lead to participants learning inconsistent associations at the level of a general food category, which would not be helpful, i.e. they may learn a 'food = 50% Go, 50% No-go' association, similar to some 'sham' training tasks.

Maximise Attention to HFSS food cues
Increase attention to HFSS food cues to enhance learning of stimulus-inhibit associations Participants attend to all images either because they need to respond to some aspects such as item location ( Figure. 1, 4; [14]), or they act as the stop cues (e.g. [9•]), or they appear shortly (100 ms) before the No-go signal (e.g. [18]). Unpredictable (50% Go and No-go) filler items (Figure 1, 3) may also help to maintain challenge and attention.
There is no need to attend to the food images, for example, the no-go cue is a simultaneously presented letter and Go trials only require a simple (e.g. space bar) response rather than indicating the item location (e.g. [48]). Note -there are no published real-world trials of this precise form of RIT for food but there are some null findings for trials for alcohol [49] or combined RIT/attention training for food [38]. explore the optimal control condition, using the other recommendations in this review to design strong trials and facilitate interpretation of findings. Trials should also measure participant blinding and expectations as these may also contribute to RIT effects [5•].

Outcome
Trials of food RIT have examined a range of outcomes (Table 1), including intake (predominantly measured using self-report via food diaries or food frequency questionnaires), weight loss (mainly measured by a researcher, but sometimes self-report) and tasks measuring putative mechanisms (e.g. devaluation, improved inhibitory control).
A limitation in the field is the over-reliance on self-reported intake as an outcome, which has generally not shown positive effects of RIT (i.e. positive in two out of eight trials [7,14]). Of note, four of the trials reporting null or mixed effects on self-reported intake showed positive effects on other outcomes [5•,7,14,18]. Reductions in selfreported intake are often observed in both the intervention and control groups, likely due to non-specific trial effects such as the recruitment of motivated participants, placebo effects and engagement in self-monitoring. Self-reported intake, particularly when measured using simple instruments such as food frequency questionnaires, is likely to be less accurate, sensitive and reliable than objectively measured intake, which has shown more consistent and robust effects of RIT in many lab studies (e.g. see [23]). In line with this, of the three trials that included an objective measure of food intake under controlled conditions, two reported positive effects of food RIT [15••,16] and one was negative [17], although the latter used an intervention that may not have trained associative inhibition (Table 1).
An advantage of RIT for food (over alcohol and smoking) is the relative ease of measuring weight loss objectively, and this has been done in eight trials. Three of these found greater weight loss in active versus control groups, at least in more at-risk populations [11••,14,18] and five reported null effects. As discussed above, most of the null trials used an intervention that may not have effectively trained associative inhibition [6,8] or recruited a suboptimal population (i.e. those who may not frequently consume the trained no-go foods [5•,19]). Memarian and colleagues [15••] attributed their null effects on BMI in children to a lack of dieting motivation or the sweetsspecific nature of their training. In order to draw clearer conclusions about RIT effects on weight loss, we need more well-designed trials in at-risk individuals.

Outcomes related to mechanisms
Several studies have reported significant devaluation of inhibited foods, measured using subjective ratings of liking, attractiveness or daily cravings [5•,7,9•,13,14,19] and fMRI cue reactivity in brain-reward systems [9•,19]. In contrast, two trials reported null findings for devaluation measured using explicit ratings [17] or an implicit association test [11••]. Explicit measures have greater sensitivity for detecting training-related devaluation [19] and the null result reported by Poppelaars and colleagues [17] may have been due to a suboptimal intervention.
In terms of improvements in inhibitory control to trained stimuli, it is important to distinguish between automatic, associative inhibition and more 'top-down' forms of inhibition (see [30]). Trials measuring top-down inhibitory control (i.e. in a separate SST or GNG, given before and after training and using different stimuli to those in the training task) have shown negative or unclear RIT effects [8,16,17]. In contrast, as described above, trials that have measured stimulus inhibition or stimulus-Go associations within the training task itself have all shown evidence of S-R learning [5•,7,13,14,16,19]. This

Current Opinion in Behavioral Sciences
Example of a RIT task containing key elements to encourage associative learning of HFSS-food inhibition (see Box 1 table above). Taken from [14] . Task adapted for use in several trials reporting positive results [5•,7,13,16,19]. It includes 1) a limited number of items/categories of HFSS food that are consistently and repeatedly paired with No-go cues.
2) A distinct category of healthy foods is always paired with Go cues and 3) a third distinct category of fillers (e.g. clothing items) is equally paired with Go and No-go cues. The fillers provide a baseline for measuring associative responses to food items, keep the task challenging and help to make the foods relatively more predictive. discrepancy between different measures of stimulus-related inhibition (measured within the training task vs. using a separate task with different stimuli) has even been demonstrated within a single study [16]. This underscores the importance of integrating and reporting measures of S-R associative learning within the training itself. Finally, the fact that those studies that did not show evidence of devaluation or improved associative inhibition also failed to find training effects on more distal outcomes further lends support to these as potential mechanisms of RIT (see also [40]).

Conclusions and future directions
This review of 16 trials of food RIT argues that positive findings are related to the use of tasks that train more associative forms of inhibition in more at-risk populations, and measure effects on more objective outcomes. It encourages researchers to measure S-R learning within their training tasks and to include simple standard outcomes (such as explicit devaluation of HFSS foods measured using visual analogue scales [32•]) to facilitate comparison between trials. Such comparison would help to address important unanswered questions such as the most effective amount and timing of RIT [46], whether training people to Go to healthy foods (in addition to No-go to HFSS foods) can increase their valuation and intake (e.g. [5•,46]) and whether training is best completed on a static computer or a mobile device [46]. By implementing the current recommendations in adequately powered, pre-registered trials, future research will be able to clarify whether, and how, food RIT translates into real-world changes in eating behaviour.

Conflict of interest statement
None