Systematic review and meta-analysis of remotely delivered interventions using self-monitoring or tailored feedback to change dietary behavior

ABSTRACT Background Self-monitoring (SM) of diet and tailored feedback (TF) have been suggested as tools for changing dietary behavior. New technologies allow users to monitor behavior remotely, potentially improving reach, adherence, and outcomes. Objective We conducted a systematic literature review and meta-analysis to address the following question: are remotely delivered standalone (i.e., no human contact) interventions that use SM or TF effective in changing eating behaviors? Design Five databases were searched in October 2016 (updated in September 2017). Only randomized controlled trials published after 1990 were included. Trials could include any adult population with no history of disordered eating which delivered an SM or TF intervention without direct contact and recorded actual dietary consumption as an outcome. Three assessors independently screened the search results. Two reviewers extracted the study characteristics, intervention details, and outcomes, and assessed risk of bias using the Cochrane tool. Results were converted to standardized mean differences and incorporated into a 3-level (individuals and outcomes nested in studies) random effects meta-analysis. Results Twenty-six studies containing 21,262 participants were identified. The majority of the studies were judged to be unclear or at high risk of bias. The meta-analysis showed dietary improvement in the intervention group compared to the control group with a standardized mean difference of 0.17 (95% CI: 0.10, 0.24; P < 0.0001). The I2 statistic for the meta-analysis was 0.77, indicating substantial heterogeneity in results. A “one study removed” sensitivity analysis showed that no single study excessively influenced the results. Conclusions Standalone interventions containing self-regulatory methods have a small but significant effect on dietary behavior, and integrating these elements could be important in future interventions. However, there was substantial variation in study results that could not be explained by the characteristics we explored, and there were risk-of-bias concerns with the majority of studies.


INTRODUCTION
A "Western diet" typically consists of intake high in saturated fats, salt, and sugars and low in fruit and vegetables, with most of the population in developed countries not meeting the WHO nutrient recommendations (1). This kind of poor diet is implicated in several chronic noncommunicable diseases (diabetes, some cancers, and cardiovascular disease) (2), and responsible in England for >10% of mortality and morbidity (3); hence, the development of public health initiatives targeting this area (4)(5)(6).
The mechanisms by which interventions can induce behavior change have been classified into 93 different techniques in the latest taxonomy iteration (7). One review of components associated with increased effectiveness in dietary and physical activity (PA) interventions found that self-regulatory behavior change techniques (SBCTs), e.g., self-monitoring (SM) and tailored feedback (TF), were associated with positive changes in dietary outcomes and differences in how interventions were delivered (provider, settings, or modality); moreover, study population characteristics did not appear to be associated with differences in effectiveness (8). Additionally, synergistic effects may occur when SBCTs are combined with other methods that target future performance, particularly those derived from control theory (9).
For diet and PA interventions, SM requires the recording of behavior, e.g., intake or activity, by an individual to actively track trends (10), primarily initiated to motivate modification of unwanted dietary or PA behavior. For SM to be successful, consistent and frequent recording is required (11), with selfevaluation being the next step, followed by self-reinforcement (12,13). To achieve the recommended behavioral change, i.e., to promote meaningful alterations and maintain wanted behaviors, individuals must analyze their actions, change them accordingly, and preferably repeat the cycle of evaluating behavior against their incorporated standards (13).
On the same continuum as SM, providing TF [where a user's unique characteristics are utilized-e.g., previous actions (14)] can be effective in changing future behavior. Compared to generic information, TF has the potential to provide more individualized information that is perceived as more salient by the user, increasing the likelihood of adhering to such advice (15,16).
We did not distinguish between interventions that use SM or TF in this review, though they differ in who is providing the results of previous behavior, because they often intertwine and hence can be difficult to separate out; indeed, under the behavior change technique classification, they come under the same cluster (7). This is especially pertinent when technology is used, as many apps contain both SM and TF as integral components, e.g., an app that provides users with a breakdown of the nutrient composition of their meal after they enter the foods they have eaten.

Aim
Our systematic review aimed to answer the question: are remotely delivered interventions that use SM or TF effective in changing eating behaviors?

METHODS
We used initial searches to identify keywords in PubMed to develop a strategy for adaptation to other databases (Embase, CENTRAL, PSYCHINFO, and Web of Science). The search, which was conducted in October 2016 and updated in September 2017, was restricted to those articles published after 1990 in peer-reviewed literature, in English, French, or Spanish. The protocol was placed in advance on PROSPERO: CRD42016042015 (http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID =CRD42016042015) and contains the search strategies used.
In order to be included, trials needed to use a remotely delivered (i.e., any standalone method that does not use direct human support) dietary SM or TF intervention in the intervention arm only and contain outcomes pertaining to most dietary consumption behaviors. We limited our study type to randomized controlled trials. Studies were excluded if they were based on populations that included children <18 y of age; those with impairments leading to disordered eating (e.g., anorexia nervosa); mixed behavior interventions (e.g., diet and exercise); interventions delivered face-to-face, in groups, or via telephone or video calls; feedback solely tailored to characteristics other than previous dietary behavior (e.g., feedback tailored by demographics); outcomes only measuring weight, calories, micronutrients (except salt), total carbohydrates, or protein (these outcomes were not included as, depending on study aims and populations of interest, these outcomes could be intended to increase, decrease, or remain the same. However, all other dietary consumption behaviors were included as outcomes for the review).
Three assessors independently screened consequent samples of 5%, calculating interrater agreement with Cohen's and Fleiss's κ. At 10%, substantial agreement had been reached (Cohen's κ = 0.62-0.87, Fleiss's κ = 0.67-0.72) (17) and the remainder was split between the assessors. Two reviewers extracted the study characteristics, intervention details, and outcomes and assessed the risk of bias using the Cochrane tool (18). Subgroup analyses were conducted stratified by risk of bias, with studies categorized as being at a high risk of bias if they were rated as a high risk for any bias included in the Cochrane tool with the exception of performance bias (due to the nature of the intervention, virtually all of the studies did not blind the participants to their intervention status).

Data synthesis
For each study, data were extracted on the first reported measures either during the intervention period or instantly after it to assess the immediate impact of the intervention. For each dietary outcome, we extracted data on the mean and SD within the control and intervention groups. Where results were reported for >1 control or intervention group, the mean and SD were combined using standard methods described in the Cochrane handbook (18). In some cases, data were extracted on the mean and SD in the change of outcome between intervention period and baseline, but only when data on mean and SD of the outcomes themselves were not available; this was not chosen as the primary outcome as it was only adequately reported in a small minority of studies. All results were converted to standardized mean differences (SMDs) using Hedge's g statistic, as described in the Cochrane handbook (18). For dietary outcomes where the intervention aims to reduce consumption (e.g., saturated fat intake), we multiplied the SMD by -1, so that in each case a positive SMD represented an improvement in diet.
All of the dietary outcomes were included in a 3-level, randomeffects meta-analysis, with both individuals and dietary outcomes nested in studies. This method explicitly accounts for correlation within studies of different dietary outcomes (19). Heterogeneity was measured using the I 2 statistic. The inclusion of multiple outcomes from single studies precludes the use of funnel plots to assess the risk of publication bias, and therefore we conducted a multilevel (outcomes nested in studies) meta-regression of the effect of standard error on effect size; in the absence of publication bias, these 2 variables should not be related. As a sensitivity analysis, we conducted a "one study removed" analysis to assess whether our meta-analysis results were overly influenced by any of the included studies. We explored reasons for heterogeneity by conducting univariate multilevel (outcomes nested in studies) regression analyses of the impact of the following variables on the effect size: risk of bias; type of dietary outcome (fruit and vegetables, fatty acids, or other); geography (Europe, United States, or other); mode of delivery of intervention (mobile phone, website, or other); length of intervention period; population type (general population, or specified by cardiovascular risk factor, e.g., overweight); and method of diet measurement (food diaries, food frequency questionnaires, or other). All analyses were conducted in R version 3.2.2, using the metafor and lme4 packages (20)(21)(22). Figure 1, after duplicates were removed, we retrieved 6838 articles. After abstract and paper screening, this was narrowed down to 27 articles (see Table 1)  reporting on 26 studies containing 37 interventions for inclusion, of which 23 studies (23,24,(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(41)(42)(43)(44)(45)(46)(47)(48)(49) were included in the metaanalysis.

FIGURE 2
Bias assessment using the Cochrane assessment of bias tool. (29,37,39). No information was provided to the control group in 4 studies (30,40,41,44). In 4 studies the intervention was provided to the control group at the end of the study (35,38,45,47); additionally, in 1, another part was given group education instead (47). The remaining 2 were given either health professional visits (46) or nonnutritional information, i.e., stress management (26).

Outcomes
Multiple dietary outcomes were collected in most of the studies included in the meta-analysis, including: total fat (n = 11)  (26,35,40,41,45,46,48). In the remaining 2 studies, 1 used both the questionnaire and the diary (47), whereas purchased food was assessed in the other (34).
Three trials were excluded from the meta-analysis as we could not extract numeric outcome data (25,26,40); 2 of these showed no change in the intake amount of either fruit and vegetables (26) or dairy (40). Atienza showed an increase of vegetables by 1.5-2.5 servings/d (P = 0.02) and in fiber by 3.7-4.5 servings/d (P = 0.10) (25). Figure 2, most studies were found to have an unclear risk of bias in the domains of random sequence generation, allocation concealment, blinding of outcome assessment, and selective reporting due to a lack of clarity about how randomization, blinding, or analysis was undertaken. No studies had a high risk of bias in either allocation concealment or blinding of outcome assessment. Due to the nature of the interventions, it is difficult to blind participants, and hence most studies were found to have a high risk of bias in the related section. If this domain is disregarded, only 5 studies did not contain high risk of bias judgments.

Meta-analysis
The multilevel meta-analysis revealed a pooled SMD of 0.17 (0.10, 0.24; P < 0.0001), indicating a significant improvement in diets as a result of tailored feedback and/or self-monitoring of diets (Figure 3). An SMD <0.2 is sometimes described as a small effect (50). The I 2 statistic for the meta-analysis was 0.77, indicating that 77% of the variance in the final result was due to between studies variance.
Multilevel regression (outcomes nested in studies) showed that the standard error was strongly positively associated with effect size, indicating potential publication bias. However, 1 study (38) was an outlier in the regression, with very large SE and effect size compared to the other results. Therefore, we conducted the regression analysis again with this study excluded. The results showed no effect of SE on effect size (β = 1.11, SE = 0.73, P = 0.131), indicating no evidence of publication bias. The "one study removed" sensitivity analysis revealed that no single study unduly influenced the results of the multilevel meta-analysis. The pooled effect size in the sensitivity analyses ranged from 0.12 to 0.18 and the P values were always <0.001. The subgroup analyses stratified by risk of bias showed very little difference in studies with low risk of bias [SMD 0.17 (0.05, 0.29), P < 0.01] and high risk of bias [SMD 0.17 (0.08, 0.26), P < 0.001] (see Supplemental Figure 1 and Supplemental Figure 2). Our multilevel regression analyses exploring reasons for the heterogeneity in results found only 1 variable where differences in effect size were significant at the P < 0.05 level. Studies that measured diet quality using means other than food diaries or food-frequency questionnaires produced larger results (P = 0.028). The only other result with borderline significance was that results for dietary outcomes other than fruit, vegetables, or fatty acids tended to be larger (P = 0.077) (see Supplemental Table 1 for full results).

DISCUSSION
Our review showed a positive but small change in diet as a result of SM or TF (based predominantly on studies of TF), although with high heterogeneity between results. That is to say, remote interventions using self-regulation methods do influence dietary change for the better and, potentially, if this effect was extrapolated over a population, it could produce a significant impact (51). This is despite potential barriers such as cost  Brug et al., 1998 (28): "The fat score that ranges between 12 and 60 is the result of a short [FFQ] in which the frequency of use and portion size of the 12 main fat sources in the Dutch diet are assessed"; Campbell et al., 1994 (29): "Dietary fat and saturated fat scores were obtained by multiplying frequency of consumption (calculated as servings per day) by portion data for each item and summing the items"; Campbell et al., 1999 (30): "Dietary fat scores were obtained by multiplying frequency of consumption adjusted to daily intake (3, 2, 1, 0.5, 0.14, 0.07 and 0) by fat content per serving of each item and summing items"; Gans et al., 2009 (31): "The FHQ fat summary score was calculated by taking the mean of all behavioral FHQ questions … response categories for the behavioral questions were: 0 = almost always, 1 = often, 2 = sometimes, 3 = rarely, and 4 = never"; Oenema et al., 2005 (39): "Answers to the [FFQ] items were converted into a fat score ranging from 0 to 80, reflecting total saturated fat intake"; Springvloet et al., 2015 (42, 43): "Saturated fat intake was measured with [an FFQ] … Based on this questionnaire, fat points were calculated … The total 'fat score' was based on 35 … food products [to which] … fat points were assigned for each product group, ranging from zero … -5 (…summed up to create a total fat points measure)." EDNP, energy dense, nutrient poor; FFQ, food-frequency questionnaire; FHQ, food habits questionnaire; Sat, saturated; serv, servings; SSB, sugar-sweetened beverage; Veg, vegetables; %kcal, percentage of total.
(52)-following dietary recommendations has been found to cost more and hence can become unaffordable amongst lower socioeconomic classes (53, 54)-as well as restrictions that participants face on time, motivation, social support, organizational demands, and emotional availability (55,56).
As this review shows that approaches not requiring instantaneous personal contact can positively impact on diet, this has implications for rolling out and integrating digital health interventions into mainstream clinical practice. Media that have a high potential reach and widespread usage, such as supermarket loyalty cards (57) and health apps-both mobile (58) and online (5, 59)-can potentially raise adherence and the measurement accuracy of SBCTs. In addition, they can be programmed to provide advice that accommodates personal needs and preferences, leading to potential improvements in associated health behaviors (56,59,60). It is easier to implement digital interventions at scale (8,56), but most of the evidence identified in this review is from studies that used more traditional research media (e.g. paper diaries).

Other literature
Although there have been previous systematic reviews that have looked at SM and TF interventions (9,10,15,(60)(61)(62)(63)(64)(65)(66)(67), our review was more inclusive on population characteristics, outcomes, and delivery methods. In contrast, prior reviews have the following limitations: 1) they have focused on those without health difficulties (9) or a narrow population such as obese participants (10); 2) they have concentrated on the intention to change (67) or weight loss (60); 3) they were restricted to particular food groups (63); 4) they used only specific delivery media (61,66); and 5) they did not discriminate between a wide variety of behavioral techniques (65) or did not isolate the dietary component of the intervention (15,62). Our work shows similar results to these systematic reviews discussed below which have shown small positive effect sizes on several discrete outcomes (9,62,64). Additionally our review extends this, through combining a broad range of dietary outcomes to show an overall effect, by using the technique of multilevel meta-analysis (19), enabling us to cope with within-study correlations between study outcomes.
The review by Broekhuizen et al. (62) looked into computer tailoring of education for nutrition and PA compared to generic or no information. In the dietary domains, a favorable significant effect was found in fat (81%), fruit and vegetables (83%), and both studies on fiber. However, this was not the case in interventions on grain, added sugar, or dairy, which were comprised of one study each.
Eyles and Mhurchu (64)  In a different analysis, healthy eating was looked at separately from PA, though SM was not examined independently but combined with the following: 1) provision of feedback on performance; 2) prompt intention formation; and 3) specific goal-setting or review of behavioral goals. In these 13 trials, the effect size grew to 0.54 (95% CI: 0.21, 0.86), in contrast to 0.24 (95% CI: 0.18, 0.29) in the remaining 40 studies.

Strengths
As mentioned above, compared with others, our review includes a broader range of population characteristics, outcomes, and intervention types despite focusing purely on the area of diet. The multilevel meta-analysis meant that we could include multiple dietary outcome results from single studies and could include results from different dietary outcomes across different studies. This greatly increased the statistical power of our meta-analysis and expanded the range of our systematic review, under the assumption that the different interventions were comparable examples of SM or TF (68). Due to the large amount of heterogeneity in the intervention designs, we conducted random-effects metaanalysis.

Limitations
As we did not consult gray literature, we may have not identified all the relevant literature. Moreover, although the studies comprised a wide array of participants, environments, and modalities, they were only performed in high-income countries and mostly using nondigital methods, and therefore caution needs to be exercised in extrapolating the conclusions. Additionally, our results may be affected by misreporting; dietary intake is well known to be underreported (69) and this is not helped by having no consensus over the questionnaire or diary used for recording outcomes.
Finally, we were not able to assess long-term effectiveness of SM and TF as we only analyzed the data at the 6-mo stage and did not consider any follow-up results. Indeed, in comparison to other methods that are not SM or TF, it has been postulated that there is no sustained change long term, and when used for weight management, a decline in adherence tends to occur after 1 mo (56). Perhaps this is due to shifting motivations from those who prompted the initial change or an increased requirement of limited self-regulatory effort as automatic habitual responses are not yet fixed (65,70).
Due to the sparse use of SM as an intervention on its own twice and together with TF in 3 trials, it is difficult to draw conclusions about its effectiveness, alone or combined. Additionally, we did not measure other intervention methods used in the trials such as goal setting (7), and as such it is hard to distinguish the individual contribution to the overall effect.

Recommendations
We noticed that no studies had fiber as an outcome, despite evidence that an increased fiber intake can have health benefits (2). There were no trials in low-or middle-income countries. Therefore, we recommend the above areas for future research. It would also be beneficial to improve the reporting of group sizes, determine the reasons for drop-outs, ascertain the blinding procedures (particularly for data analysis) and consider how missing data have been dealt with so that discerning bias is easier. It will also be important to develop strategies to overcome the obstacles to using SM and TF, such as declining motivation, that have been identified.

Conclusion
This systematic review of the literature set out to answer the question: are remote interventions that use SM or TF effective in changing eating behaviors? The meta-analysis showed a significant-albeit small, heterogeneous, and at risk of biaspositive effect of standalone SM and TF on dietary change (based predominantly on TF studies in this review). At a population level, such an intervention could have an appreciable impact. Given the ability of digital health interventions to deliver interventions remotely and to reach wide audiences (6), such interventions have the potential to make a contribution to improving the healthiness of diets.
We thank Nia Roberts for her assistance with developing the initial search strategies for this systematic review. The authors's responsibilities were as follows: NT, CP, GC, and PS: designed the study; NT, AE, FB, CP, GC, JHB, and PS: developed the protocol; NT, AE, and FB: conducted the searches; NT and RS: conducted the data extraction; NT, RS, and PS: conducted the analyses; NT and PS: wrote the first draft of the manuscript; and all authors: read and approved the final manuscript. None of the authors have any conflicts of interest related to this study.