Improving body image at scale among Brazilian adolescents: study protocol for the co-creation and randomised trial evaluation of a chatbot intervention

With a population of over 212 million people ([7]), body image concerns in Brazil are common across the lifespan, particularly among adolescents. Prevalence of body image concerns range between 26.6–56% and 10.7–36% for adolescent girls and boys, respectively [8‐10]. Concurrently, up to 10% of adolescents report engaging in disordered eating and unhealthy weight control behaviours [11, 12]. Further, consumer trends show Brazil is the leading country in diet pill consumption and cosmetic surgery, with 2.5 million surgical and non-surgical aesthetic procedures conducted annually [13]. This is concerning given that body image is a key risk factor for depressive symptoms, smoking, substance misuse and self-harm in other countries [14]. Further, cosmetic procedures incur significant financial expense and have potential medical complications [15]. Together, these findings highlight the urgency for easily-accessible, scalable, and cost-effective interventions designed to address body image concerns among Brazilian populations. At present, such interventions are limited.

Existing body image interventions in Brazil have evolved from international collaborations between Brazilian and US researchers. These include the translation and adaptation of a group-based cognitive dissonance intervention, The Body Project, which provides a forum for girls and women to challenge unrealistic appearance ideals and develop body satisfaction. The program proved effective at improving body image, negative affect and resilience to sociocultural pressures, but not disordered eating, among at risk Brazilian adolescent girls [16]. New Moves, a US program, was also adapted for use with Brazilian adolescents [17]. This school-based program used a combination of motivational interviewing and exercise and nutrition sessions to improve body dissatisfaction, self-esteem and healthy eating behaviours. Despite its efficacy in improving these outcomes among US girls, it did not lead to significant improvements in these key outcomes among Brazilian girls.

To our knowledge, no evidence-based body image intervention has been developed specifically with and for Brazilian adolescents, nor has a research group utilised technology and multi-disciplinary partnerships to develop, evaluate, disseminate and implement an intervention in Brazil. First, given that 24.3 million young people aged 9–17 years in Brazil (86% of the age range) use the internet [18], digital interventions are likely to be accepted by young people, as well as highly amenable to their lifestyles, which are increasingly digitally oriented. Second, the diverse knowledge and expertise held by the partnership stakeholders ensures that relevant insider knowledge from the respective markets are integrated into the intervention development, evaluation, implementation and dissemination frameworks; thus, making for an innovative and well-considered intervention.

The emergence of digital interventions has increased the accessibility, scalability, and cost-effectiveness of mental health interventions [5, 6, 19]. A review into the adaptation of face-to-face body image and eating disorder prevention interventions (e.g., psychoeducation, media literacy and cognitive dissonance) for online use show promising results. However, effects are small and significant issues with attrition and adherence are noted (Cohen’s ds .24–.42) [20, 21]. Consistent with recommendations for using disruptive innovation (e.g., ‘edutainment’; educational content within entertainment settings) and technology to scale up mental health interventions [22], researchers have begun developing and testing purpose-built digital body image interventions. Many of these interventions have been designed to be brief, standalone approaches that provide immediate (‘in-the-moment’) symptom reprieve; an intervention model referred to as ‘micro-interventions’ (e.g., audio and visual clips, online games and e-books) [23]. An advantage of micro-interventions is the ability to embed theoretically driven, evidence-based therapeutic techniques into existing digital platforms and social media networks already frequented by adolescents; thus, potentially removing barriers to dissemination and uptake [5, 6, 22].

Micro-interventions are a promising strategy for addressing mental health among Brazilian adolescents, particularly due to the high percentage of young people with internet access (86%) and a social media network (68% [18];). Recently, micro-interventions have proven effective at producing immediate and short-term body image and mental health benefits among adolescents and young adults in Australia [23], the UK [24] and the US [25]. Until now, chatbot technology has not been developed or evaluated within a micro-intervention framework [26], nor has an evidence-based micro-intervention been purpose-built for Brazilian populations. However, a recent public health initiative by UNICEF Brazil [27], reported high uptake and acceptability of a chatbot addressing sexting among adolescents, with over one million users in the first 12-months of dissemination.

Chatbot technology is a robust, low-barrier alternative to traditional mental health approaches, which are largely dependent on health professionals and services for effective implementation and dissemination [5, 6]. Chatbots are usually hosted on a website or app, with the centralised platform allowing users to engage with a variety of intervention techniques, monitor their symptoms and progress, and at times refer at-risk individuals to more appropriate mental health resources and services [26]. This independence from traditional health care approaches also allows for around the clock access and care, particularly to those who have limited mobility, live distant from health professionals and services, or are reluctant to seek mental health advice due to stigmatization [26]. Therefore, given the prevalence of body image concerns among adolescents, coupled with the scarcity of extant prevention and interventions approaches, as well as adolescent’s frequent online behaviours, chatbot technology offers a new and innovative approach to addressing these concerns at-scale.

This paper outlines the protocol for the co-creation and controlled evaluation of a chatbot targeting risk and protective factors for body image concerns among Brazilian adolescents. A two-arm fully remote randomised controlled trial (Fig. 1) will compare a chatbot intervention to an assessment-only control condition. Participants in the intervention condition will interact with the chatbot over a 72-h period, while participants in the control condition will receive ‘standard care’ for body image concerns among adolescents in Brazil, which at present is no intervention. This protocol paper and the randomised controlled trial design were guided by the Standard Protocol Items: Recommendations for Interventional Trials (SPIRIT; see Fig. 2 and supplementary materials) and the Consolidated Standards of Reporting Trials (CONSORT [28];), including the extension for Electronic Health (CONSORT - EHEALTH [29] see supplementary guidelines).

Assessment of primary and secondary trait outcomes will be conducted at baseline (T1), post-intervention (i.e., 3 days; T2), one-week (T3) and one-month follow-up (T4). Participants in the intervention condition will also be assessed on state-based body satisfaction and affect immediately after engaging with each technique. To our knowledge, this is the first chatbot to target body image among Brazilian adolescents. Therefore, a secondary aim is to assess users’ engagement and adherence. Hypotheses were formulated for overall chatbot efficacy (H1-H2), moderation effects within sub-samples (H3), and intervention engagement and adherence (H4):

Research measures are presented in Table 3. One measure was purpose-built for the current research (i.e., body image self-efficacy). Acceptability and adherence measures were informed by existing reviews [52, 53], with remaining measures validated for use among adolescents.

A community sample of 2800 adolescents aged 13–18 from diverse ethnic, geographic and socio-economic backgrounds will be recruited via a Brazilian research agency’s recruitment panels (i.e., via email to their participant databases) and via UNICEF’s online communication platforms (e.g., U-Report¹). Recruitment will occur 2 weeks prior to the trial commencing. Given UNICEF’s reach and expertise with underserved adolescents in Brazil, the inclusion of this recruitment channel will ensure that a diverse sample is approached for participation. Further, recruitment numbers across these metrics (e.g., geography) will be regularly monitored and addressed by the research team to ensure diversity across the sample is achieved. Recruitment materials will include a link to an online screener questionnaire to determine participant eligibility. After being screened as eligible, participants will be emailed an information pack to be read with their parent or guardian. Parents will provide consent as per the Brazilian General Data Protection Law, by uploading their signature to online identification verification software that is monitored by the research agency. Following parental consent, adolescents will complete the online self-report assessment (T1), which will commence with adolescents providing informed assent. Those who do no assent will not proceed with the trial. On completion of the survey, adolescents will be randomised into one of two conditions: intervention or assessment-only. Those in the intervention condition will receive a link to the chatbot 24-hours after completing the baseline assessment. Those in the assessment-only condition will be provided with a link to the chatbot after they have completed all assessment timepoints.

Adolescents in the intervention condition will be encouraged to engage with the chatbot as much as possible over a 72-hour period. During this time, participants may complete the eight intervention techniques as many times as they like. Participants will be assessed on state body satisfaction and affect when they first enter the chatbot and immediately after engaging with a technique. If participants do not engage with the chatbot after 12-, 16 and 23.5 hours, they will receive a prompting notification that encourages engagement with the chatbot. That is, if they do not engage after the 12-hour prompt, they will receive up to two more prompts.

At completion of the 72-hour intervention period, both conditions will be assessed on primary and secondary trait outcomes immediately post-intervention and, at one-week and one-month follow-up. Those in the intervention condition will be encouraged to continue using the chatbot. Following the one-month follow-up assessment, all participants will be debriefed on study aims and objectives, and provided with mental health resources accessible in Brazil. Assessment-only participants will be invited to engage with the chatbot; however, their engagement will not be monitored or assessed. Lastly, to compensate participants for their data usage, they will receive an electronic voucher to the value of R$100 and R$80, for the intervention and control condition, respectively. Participants were not notified of the compensation value until study completion. 

In the present trial, adolescents will be randomised to one of two conditions based on a permuted block design, using survey software by the research agency. Randomisation will be performed on an individual level to ensure timely allocation, while maintaining a balance of participants across conditions. Participants cannot be blinded to allocation due to the behavioural nature of the two conditions; however, they will be blinded to the study aims and hypotheses. As randomisation is conducted by an external research agency, researchers will be blinded to participant allocation. Data analysts will be blinded to condition allocation (dummy coded) when conducting statistical analyses on primary and secondary trait outcomes; however, blind analyses cannot be conducted on primary and secondary state measures due to the within-group design. State and trait data will be separated into two data files to ensure blinding is maintained for trait-based measures.

The limited number of prior micro-interventions and insufficient detail for key parameters complicates efforts to confidently generate all required parameter estimates for power calculations. Accordingly, we adopted a Monte Carlo simulation approach that tested power across a range of combinations of key statistical parameters for multilevel modelling: within-person sample size (varying from 3 to 5 estimates per person), intraclass correlation (ICC) for outcome (ranging from .3 to .6 in recognition that ICC values tend to be large for repeated measures designs), random slope variance (choosing .01, .09, and .25 to reflect small, moderate, and large amounts of variance [54]), and effect size (setting at r = .2 to reflect small yet practically meaningful effect size [55]).

Using the SIMR (simulation-based power calculations for mixed models) package in R [56], we tested power across these different combinations of parameter estimates, with power set at .80, alpha = .05 (two-tailed), equal allocation to intervention and control arms, and conservatively allowing for 25% attrition on an initial sample of 2800 participants at the final timepoint. We note too that Hypotheses 1 and 4 are limited to the intervention arm, and thus the sample size used for those calculations was 1050 (1400 participants with 25% attrition). Using these inputs, our simulations indicated power was > .99 for all cross-level interaction effects regardless of input values, and > .80 for all within-person effects with the exception of scenarios where: (1) the random slope variance was small, or (2) a small-to-moderate random slope variance was accompanied by a small number of within-person estimates (average cluster size of 3–4). These scenarios of small random variance coupled with a small estimated effect size are unlikely based on best available evidence from past studies [23‐25]. Thus, our sample size is likely to be sufficient across a range of plausible data scenarios.

At trial completion, Toluna and Talk2U will deliver password protected datafiles to the corresponding author who will clean, screen and separate state- and trait-based data for blinding purposes. Data will then be delivered to the data analysts. Analyses will be conducted in Stata. A series of linear mixed effects models will test hypotheses. H1 analyses will be limited to intervention arm participants as there is no comparable pre- and post-technique data for the assessment-only arm, while analyses for H2, H3 and H4 will include participants from both study arms.

To test H1, the immediate effects of intervention content on state-based outcomes will be explored by regressing state-based outcome measures (body satisfaction and affect) onto a time variable (0 = pre-content, 1 = post-content). For H2-H4, scores on trait-level outcome variables will be regressed onto a dummy coded time variable, with baseline as the reference category compared against post-intervention, one-week, and one-month follow-up timepoints, using random effects to allow for individual differences in amount of change in symptoms over time. Group (control = 0, intervention = 1) will be included as a cross-level moderator to test group differences in change over time (H2). The moderating effect of baseline variables on this group effect will be conducted to test H3. In evaluating H4, adherence and engagement measures will be tested as moderators of change in trait-based outcome measures over time (relative to baseline) for intervention participants. Relations among random effects of Level 1 time variables will be modelled with an unstructured covariance matrix.

Baseline differences between intervention and control participants in terms of demographic factors will be tested. Any variables shown to differ across groups will be included as covariates in analyses. Linear mixed model results will be presented in adjusted and unadjusted forms to demonstrate impact of these covariates on results.

Analyses will be conducted on an intention-to-treat basis, with participant data included based on group assignment at baseline. As the post-technique state-based measure is an indicator of use of the intervention resources, only instances where participants provide data for both pre- and post-technique will be included in analyses for H1. Missing data for group comparisons (H2) and baseline moderation effects (H3) will be handled using multiple imputation by chained equations with 50 imputations. Pattern mixture modelling using the mimix package [57] will be applied to determine sensitivity of results to non-ignorable missingness.

The research will be conducted in accordance with the ethical standards and guidelines for conducting research with young people in Brazil. Study participation is voluntary. Prior to consent and assent, parents and participants will be provided with an information sheet outlining the research procedures and associated risks. Consent withdrawal is possible at any time without cause for justification. At study completion participants will be provided with a debrief form disclosing research aims and ancillary mental health sources.

The associated ethics committees will be informed of any severe adverse events or other unintended effects of the trial intervention. This study received ethics approval from the Instituto Federal de Educação, Ciência e Tecnologia do Sudeste de Minas Gerais (4.232.804), Comissão Nacional de Ética em Pesquisa (4.582.466), and the University of the West of England (HAS 19.12.090). The study is registered with Clinical Trials. Gov (NCT04825184; website: https://​clinicaltrials.​gov/​show/​NCT04825184).

Research findings will be disseminated via open-access peer reviewed publications and academic conferences, as well as beyond the scientific community. Findings will be shared through infographics and videos to interested members of the public, including on the social media platforms and websites of the Dove Self-Esteem Program, UNICEF and Talk2U. If the intervention proves effective, the chatbot will be scaled up for delivery to adolescents across Brazil via social media marketing campaigns targeting audiences directly on Facebook, Instagram and TikTok.

The current research makes several advancements in addressing mental health demands of young people at-scale and builds upon previous work in innovative ways. First, our international multi-disciplinary partnership between academics, industry organisations and UNICEF provide the expertise and infrastructure to develop, evaluate and disseminate a highly scalable evidence-based intervention for adolescents. Second, collaborative decision-making processes between adolescents and project partners, as well as a small internal pilot ensured the chatbot was highly acceptable among the target audience; a key moderator of intervention efficacy [52]. Third, the chatbot is hosted on Facebook Messenger, a free and widely available online platform. Hence, intervention implementation and dissemination are independent from health professionals, services and physical locations (e.g., schools, clinics), and thus reduce barriers to health care and increase engagement among hard-to-reach populations [26]. Relatedly, a micro-intervention framework was used to develop and evaluate the chatbot; an intervention model yet to be utilised among Brazilian adolescents. Findings from this research will further our understanding of micro-intervention effectiveness at providing immediate and short-term mental health reprieve. Lastly, body image concerns are a global issue among adolescents. Therefore, if efficacious, the chatbot may be translated and adapted for use among adolescents in other cultures and countries; thus, further expanding the reach and impact of the current research.

The current study will not be without limitations, including the use of a fully remote, online randomised controlled trial design and an assessment-only control condition. Reviews into online interventions indicate that this intervention modality is associated with low adherence and high attrition rates in trials [20, 26]. These factors were considered during intervention development and research design phases. Intervention features such as check-in messages, multi-media stimuli (i.e., audio and video clips; emojis [a small digital image or icon used to express an idea or emotion]), virtual hosts and the gamification of intervention techniques were incorporated to enhance user engagement, experience and acceptability. Further, the study is sufficiently powered to detect effects within intention-to-treat and completer samples. Recruitment and data collection will be outsourced to an internationally recognised research agency who specialise in large-scale data collection; thus, ensuring the randomised controlled trial is conducted in a timely and effective manner with a nationally representative sample. Lastly, an assessment-only control condition was informed by a public health framework for assessing interventions in real world settings, and comparing them to ‘care as usual’ [58]. Given the scarcity of body image interventions for adolescents, receipt of no intervention was deemed appropriate.