Category: 3. Business

Worldline (ENXTPA:WLN) has drawn renewed interest after launching its Android SmartPOS solution in the UK. The new platform aims to simplify payments and customer engagement for small and medium-sized businesses by unifying core operations.

See our latest analysis for Worldline.

Worldline’s push into Android SmartPOS comes as the company navigates a rocky stretch in the market. The past month has seen a 33.9% drop in its share price, while the total shareholder return over the last year has plunged 77%. Despite recent innovation, momentum remains weak and investors are waiting for signs of a real turnaround before confidence returns.

If the evolving landscape in digital payments has you thinking beyond the usual picks, now is a great moment to discover fast growing stocks with high insider ownership

So with shares trading well below analyst targets and a new product on the scene, is Worldline an overlooked value play, or is the market factoring in all the risks around its recovery and future growth?

With Worldline’s last close at €1.53, the most popular narrative sees fair value at €2.51, a sizable disconnect that is fueling debate about the company’s turnaround efforts and what’s needed to close the gap.

The launch of next-generation digital payment products (Wero in Germany, France, Belgium; refactored e-commerce platform rolled out with Credit Agricole; UK post-Brexit offering) coupled with investment reallocations from MeTS divestment supports Worldline’s capacity to capture greater volumes from the ongoing shift to cashless payments and e-commerce growth (especially in underpenetrated European and emerging markets), positively impacting future revenue and top-line growth.

Read the complete narrative.

Want to see what sits under this bold valuation call? The secret sauce is in aggressive profit margin forecasts and a projected earnings turnaround fit for a comeback story. Ready to see which financial levers carry the most weight in this optimistic scenario? The full narrative has the details the market is watching.

Result: Fair Value of €2.51 (UNDERVALUED)

Have a read of the narrative in full and understand what’s behind the forecasts.

However, ongoing declines in revenue and persistent pressure on profit margins could undermine even the most optimistic forecasts for Worldline’s turnaround story.

Find out about the key risks to this Worldline narrative.

If you see things differently or are keen to chart your own course, it takes just a few minutes to shape your own perspective. Do it your way.

A great starting point for your Worldline research is our analysis highlighting 3 key rewards and 1 important warning sign that could impact your investment decision.

Smart investors never stick to just one story. Get ahead of the crowd by checking out high-potential stocks and trends everyone will be talking about tomorrow.

This article by Simply Wall St is general in nature. We provide commentary based on historical data and analyst forecasts only using an unbiased methodology and our articles are not intended to be financial advice. It does not constitute a recommendation to buy or sell any stock, and does not take account of your objectives, or your financial situation. We aim to bring you long-term focused analysis driven by fundamental data. Note that our analysis may not factor in the latest price-sensitive company announcements or qualitative material. Simply Wall St has no position in any stocks mentioned.

Companies discussed in this article include WLN.PA.

Have feedback on this article? Concerned about the content? Get in touch with us directly. Alternatively, email editorial-team@simplywallst.com

Introduction

Background

Adolescent mental health is a cornerstone of national well-being and a pivotal focus of China’s “Healthy China” initiative. This commitment is demonstrated by consecutive national policies explicitly targeting the enhancement of mental health literacy (MHL). Key actions include the “Healthy China Action (2019-2030),” which sets specific literacy targets, and the “Action Plan for Children and Adolescent Mental Health (2019-2022),” which aimed to boost mental health knowledge among youths [,]. This policy momentum has been sustained through recent strategic outlines, underscoring the ongoing priority of safeguarding adolescent psychological well-being [,].

Against this policy backdrop, a pressing public health challenge has emerged. National data indicate that 14.8% of adolescents in China face depression risk, with 40% frequently experiencing loneliness [,]. This psychological vulnerability coexists with a near-saturated digital environment, characterized by 96.8% internet penetration among minors []. Within this context, a pronounced reliance on digital technology is evident, as over one-third of adolescents exhibit psychological dependence on smartphones. Moreover, prolonged social media use is linked to a significantly increased risk of anxiety and depression []. This combination of factors means today’s teenagers, who are “digital natives,” face 2 major risks: heightened psychological distress and problematic social media use. Therefore, improving health literacy—particularly in the digital mental health domain (mental eHealth literacy [MeHL])—and mitigating social media addiction (SMA) has emerged as a critical public health imperative.

Enhancing health literacy is widely established as a cornerstone strategy for public health improvement [,]. Yet, adolescents often exhibit insufficient health literacy, which manifests as help-seeking dilemmas (eg, “not recognizing illness” or “recognizing but not seeking treatment”) that hinder access to care []. This is particularly critical for senior high school students, whose developing literacy can shape lifelong health trajectories. In theory, health literacy is thought not only to directly affect health but also to work by influencing other factors (as a mediator) or by changing the strength of a relationship (as a moderator) [-].

However, research has seldom examined the distinct and potentially synergistic roles of its multidimensional facets in the digital context. Specifically, a significant gap remains in understanding the interplay between MHL—focused on psychological cognition, eHealth literacy (eHL)—centered on digital competency, and their integration as MeHL in relation to adolescent SMA and depression-anxiety-stress (DASS). To address this gap, this study aims to investigate the multifaceted effects and underlying mechanisms of this multidimensional health literacy framework, thereby informing targeted mental health interventions.

Theory and Research Hypotheses

Health Literacy, SMA, and DASS

Health literacy, initially proposed by Simonds in 1974, emphasizes improving public health decision-making skills through education []. Later, Nutbeam [] further expanded the concept, proposing a 3-level model of functional, interactive, and critical health literacy, focusing on individuals’ ability to acquire health information while emphasizing the ability to assess information and implement health decisions. The World Health Organization defines health literacy as “the ability to access, understand, evaluate, and apply health information and services to promote and maintain good health” []. In the “Healthy China Action (2019-2030)” [], health literacy is viewed as the ability to acquire, process, and understand health-related information, which is essential for maintaining or improving health and quality of life, and is a critical means for improving health outcomes and promoting healthy behaviors. Thus, health literacy is not limited to the acquisition of health knowledge but emphasizes the promotion of overall health through effective health decision-making and behavior change.

As an important branch of health literacy, mental health literacy was first introduced by Jorm et al [], focusing initially on individuals’ ability to recognize, understand, and address mental health issues, while emphasizing the reduction of stigma and the enhancement of help-seeking willingness []. As research deepened, the scope expanded to include multiple dimensions such as recognizing mental health problems, understanding causes, and coping strategies []. The core feature of MHL focuses on psychological content, emphasizing cognitive processing and attitude transformation toward mental states.

With the widespread penetration of digital technologies, eHL and MeHL have emerged. The former, defined by Norman and Skinner [], emphasizes the ability to search, evaluate, and apply broad health information through electronic media, highlighting the operational characteristics of media approaches. The latter, as an emerging concept, is defined as “the knowledge, beliefs, and skills to manage mental health through digital tools” [], which represents an intersection of MHL and eHL within a specific technological context. It is a dual-core integration dimension of “media × psychology.” Although these 3 aspects fall under the health literacy system, they are interconnected but not hierarchically subordinate, and should be understood as parallel and intersecting multidimensional variables []. In conclusion, health literacy is a multidimensional, systematic concept that encompasses the acquisition, understanding, judgment, and application of health knowledge and services. It is a learned psychological tendency that is highly interventionable and modifiable [].

In the current educational context, stress has become a significant psychosocial factor affecting adolescents, with anxiety and depression being the most common mental health disorders []. According to data from the China Youth Research Center [], the detection rate of anxiety among primary and secondary school students is 31.3%, with a depression risk rate of 17.9%. Additionally, 57.4% of high school students report high levels of academic pressure. Many adolescents delay or are unwilling to seek psychological support due to the stigmatization of mental health issues, lack of knowledge about mental health assistance, and insufficient external support [], reflecting a deficiency in MHL. Adolescent mental health is directly influenced by MHL. Good MHL not only helps to recognize mental disorders [] but also increases help-seeking willingness and coping ability [], alleviates negative emotions, and reduces the risk of depression and anxiety [,], with a particularly significant improvement observed in adolescents with mild to moderate depression.

eHL, which is the ability to obtain and apply online health information, is also crucial for adolescent mental health. Lower levels of eHL lead to difficulties in understanding information, which in turn affects health management behaviors. Higher levels of eHL are linked to adolescents’ ability to acquire effective health resources, higher psychological well-being, and better self-regulation abilities []. Furthermore, eHL is negatively correlated with anxiety and depression [], playing a positive role in relieving psychological stress, enhancing social support, and improving emotional adaptability [,].

SMA, as a typical behavioral addiction of the digital age, is characterized by maladaptive dependence on social media platforms and a loss of self-regulation []. A representative survey of nearly 6000 adolescents revealed that 4.5% of them are at risk for SMA [], with prevalence rates generally higher than those in university students and community adults []. Existing research primarily explains the mechanisms of SMA from the perspectives of personality traits (eg, emotional stability) and negative emotions (eg, depression and anxiety) [], with limited attention given to the role of health literacy.

Social cognitive theory (SCT) suggests that the interaction between individuals, behaviors, and environments shapes behavioral patterns, and adolescents’ social media use is influenced not only by the media environment but also by their level of health literacy []. Research has found that higher MHL is significantly associated with a lower likelihood of SMA and can effectively intervene in maladaptive usage behaviors []. eHL has also been shown to be strongly negatively correlated with internet addiction []. Adolescents with lower health literacy are more likely to develop internet dependence, using the internet to distract themselves, escape reality, or regulate negative emotions such as loneliness, anxiety, and depression []. The information overload and social pressure in the online environment further amplify their negative emotional experiences, particularly in groups with insufficient eHL, which increases the likelihood of feelings of anxiety, isolation, and self-denial [].

It is crucial to distinguish between the core constructs of “digital literacy” (eg, eHL) and the contextual challenge of “information overload.” Digital literacy refers to an individual’s subjective capability to process digital health information, acting as a protective factor. In contrast, information overload represents an objective environmental stressor or risk factor. The core logic of our framework is to investigate how high levels of digital literacy enable adolescents to effectively navigate and mitigate risks within an information-overloaded environment. Additionally, adolescents who use social media for more than 2 hours a day often experience increased DASS and suicidal ideation, highlighting the real threat of addictive behaviors. The higher the level of eHL, the more effectively individuals can use digital resources to regulate emotions and improve behavioral health, demonstrating lower addiction tendencies and higher psychological resilience.

The Impact of Health Literacy on SMA and DASS

MeHL is a composite literacy formed by integrating the core elements of MHL and eHL in the context of digital media. It emphasizes an individual’s ability to acquire, understand, assess, and apply mental health information using digital tools []. Compared to MHL, which focuses on the recognition and management of mental health problems, and eHL, which emphasizes the acquisition and assessment of digital health information, MeHL involves both cognitive processing of psychological content and the technical processing of media information, forming an integrated, rather than subordinate, independent dimension [].

As an important intermediary pathway through which education levels influence health status, lower MeHL is significantly associated with an increased risk of mental health disorders []. Individuals with higher levels of eHL and MHL are better equipped to use digital platforms to access psychological support and intervention resources. This results in stronger MeHL, which enables them to more effectively identify and address psychological risks, thus reducing maladaptive and addictive behaviors associated with social media use []. Some studies also suggest that digital literacy is negatively correlated with behavioral addiction, meaning that the higher the digital skills, the better adolescents can avoid excessive social media use [].

Further research has found that both MHL and eHL have a significant positive effect on predicting health behaviors, and that MHL can play a key moderating role in the process by which eHL influences health behaviors, with its moderating effect reaching 31.1% []. Individuals with higher levels of eHL possess the ability to filter information and operate media tools, but without the cognitive support of MHL, they may find it difficult to convert information into effective behavioral intentions. On the other hand, individuals with higher MHL are more likely to proactively identify, understand, and integrate health information in complex digital environments, thereby enhancing their behavioral execution and psychological regulation []. This mechanism aligns with the behavioral-driving logic of “perceived benefits” and “self-efficacy” in the health belief model (HBM) [] and is consistent with the “individual-behavior-environment” triadic interaction pathway proposed by SCT [].

This Study

Existing research primarily focuses on the independent roles of eHL and MHL, lacking a comprehensive examination of their interactive effects, particularly in the context of SMA. While MeHL is an emerging concept, its role and interactions within the digital environment remain underexplored. To address these gaps through an integrated theoretical lens, this study synthesizes conservation of resources (COR) theory, HBM, and SCT. These theories offer complementary explanations: COR theory provides the motivational foundation by framing health literacy as a key personal resource; HBM elucidates the cognitive decision-making process; and SCT offers the overarching behavioral framework of triadic reciprocity. Guided by this framework, we construct a comprehensive model () to elucidate the joint influence pathways of eHL, MHL, and MeHL on adolescent SMA and DASS. The following hypotheses are proposed:

H1: eHL and MHL have a significant negative effect on adolescent SMA and DASS.

H2: MeHL mediates the effect of eHL and MHL on adolescent SMA and DASS.

H3: MHL significantly moderates the effect of eHL on MeHL and, through a dual moderation mediation pathway, influences SMA and DASS.

Methods

Study Design

This study used a cross-sectional design with a stratified cluster random sampling method to examine the relationships between adolescents’ multidimensional health literacy and their engagement with SMA, as well as symptoms of depression, anxiety, and stress. The study was reported in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement [].

Setting

This cross-sectional study was conducted in 5 provinces of China (Beijing, Zhejiang, Shanxi, Henan, and Jiangsu), selected to represent a range of regional digitalization levels. Data collection was carried out in the computer labs or classrooms of the participating schools over a 3-month period, from February to April 2025.

Participants

The sampling procedure was conducted in 3 stages: first, 5 provinces were selected as strata based on regional digitalization indices; second, 2 general high schools were randomly selected within each province; and finally, 2 intact classes were randomly chosen from each school, resulting in a total of 20 classes. The target population comprised adolescents attending school, with exclusion criteria including severe physical or mental disabilities, inability to independently complete the questionnaire, or ongoing related treatment. A total of 855 adolescents completed valid questionnaires (response rate 85.5%), with 449 (52.5%) males, 406 (47.5%) females, and a mean age of 16.38 (SD 0.86) years.

Variables

This study examined the following variable classes, all derived from self-report questionnaires:

• Outcome variables: SMA and a composite score of DASS served as the primary outcomes.
• Exposure variables: eHL and MHL were the core independent variables.
• Mediating variable: MeHL was tested as the mediator in the hypothesized pathways.
• Effect modifier: MHL was additionally examined as a moderator of the relationship between eHL and MeHL.
• Covariates: Participant age and gender were included as control variables in all multivariable analyses.

Measurement

eHealth Literacy Scale

The eHealth Literacy Scale (eHEALS) developed by Norman and Skinner [] was used. This scale contains 8 items and covers 3 dimensions: the ability to acquire, evaluate, and apply health information online. It assesses adolescents’ knowledge, comfort, and self-perceived ability to acquire, understand, assess, and apply electronic health information. The scale uses a 5-point Likert scale, with responses ranging from 1 (“strongly disagree”) to 5 (“strongly agree”). The total score ranges from 8 to 40, with higher scores indicating higher levels of eHL. In this study, the scale demonstrated excellent internal consistency, with both a Cronbach α of 0.944 and a McDonald ω of 0.959 for the total score. The consistency estimates for the 3 dimensions were also high: application ability (α=0.907; ω=0.948), evaluation ability (α=0.883; ω=0.934), and decision-making ability (α=0.809; ω=0.922).

Mental Health Literacy Scale

MHL was assessed using a 12-item short form adapted for this study from the original scale by O’Connor and Casey []. To ensure brevity, we selected items that best represented the 4 theoretical dimensions: problem recognition, willingness to seek help, knowledge understanding, and attitudes toward professional help. The measure uses a 5-point Likert scale (“1=strongly disagree” to “5=strongly agree”), and a sum score is calculated (range 12-60), with higher scores reflecting higher MHL. The validity and reliability of this adapted form were supported empirically in our sample: its structural validity was upheld by confirmatory factor analysis (CFA). It demonstrated high internal consistency, with a Cronbach α of 0.946 and a McDonald ω of 0.964 for the total score. The coefficients for the 4 dimensions were problem recognition (α=0.885; ω=0.942), help-seeking (α=0.838; ω=0.919), knowledge understanding (α=0.867; ω=0.941), and attitude (α=0.901; ω=0.949).

Mental eHealth Literacy Scale

The Mental eHealth Literacy Scale (MeHLS) developed by Richard [] was used. This scale consists of 23 items, divided into 3 dimensions: cognitive (8 items), knowledge (7 items), and skills (8 items), to assess an individual’s knowledge, beliefs, and online information processing abilities in managing mental health. The scale uses a 5-point Likert scale (“1=never” to “5=always”), with a total score range of 23 to 115. Higher scores indicate higher levels of digital MHL. In this study, the internal consistency was excellent, with a Cronbach α of 0.953 and a McDonald ω of 0.960 for the total scale. The values for the 3 dimensions were cognitive (α=0.933; ω=0.949), knowledge (α=0.929; ω=0.943), and skills (α=0.937; ω=0.948).

Social Media Addiction Scale

The Bergen Social Media Addiction Scale (BSMAS), developed by Banyai [], was used for assessment. This scale consists of 6 items reflecting 6 typical addiction characteristics: salience, emotional regulation, tolerance, withdrawal symptoms, conflict, and relapse [,]. In line with its widespread application and robust psychometric validation across different cultures and adolescent samples, the BSMAS is consistently treated as a unidimensional scale, and the sum of all items is used to indicate the severity of SMA tendencies [-]. It is used to identify adolescents’ problematic social media use behaviors. The scale uses a 5-point Likert scale (“1=never” to “5=always”), with a total score range of 6 to 30. Higher scores indicate stronger tendencies toward SMA. In this study, the internal consistency for the scale was good, as indicated by both Cronbach α (0.894) and McDonald ω (0.928).

Depression-Anxiety-Stress Scale

The Depression Anxiety Stress Scale-21 (DASS-21), developed by Lovibond in 1995 and later revised and optimized by Antony, was used. This scale contains 21 items divided into 3 subscales—depression, anxiety, and stress—to assess the severity of emotional symptoms over the past week. The scale uses a 4-point Likert scale (“0=never” to “3=always”), with higher scores on each subscale indicating more severe symptoms. For the purpose of this study, a composite psychological distress total score was computed by summing the scores from the three DASS-21 subscales (depression, anxiety, and stress). For the primary analyses (including correlation and mediation analysis), a composite “psychological distress” total score was computed by summing the scores of the 3 subscales. The scale demonstrated excellent reliability in our sample. The internal consistency estimates for the subscales were as follows: depression (Cronbach α=0.909; McDonald ω=0.946), anxiety (α=0.898; ω=0.939), and stress (α=0.914; ω=0.947). The psychological distress total score also showed excellent internal consistency (α=0.959; ω=0.963).

The scales used in this study were all based on mature tools developed both domestically and internationally, or were revised and localized based on previous research. They are grounded in clear theoretical frameworks, scientifically structured designs, and supported by solid practical applications. The full survey questionnaire is available in . To assess the construct validity of the scales, CFA was conducted to evaluate the goodness-of-fit of each measurement model.

The results showed that all models met the evaluation criteria: the χ²/df ratio was within a reasonable range, and the robust comparative fit indices (comparative fit index, Tucker-Lewis index, and incremental fit index) consistently exceeded the recommended threshold of 0.90, with most surpassing 0.95. While the goodness-of-fit index and adjusted goodness-of-fit index for the more complex DASS-21 model were slightly below 0.90—a common occurrence for multidimensional scales—the collective evidence strongly supports model fit. The root-mean-square error of approximation was less than 0.05, and the factor loadings were all at moderate to high levels, indicating that the scales have good construct validity. In conclusion, the measurement tools used in this study meet psychometric standards in terms of reliability and validity, ensuring high measurement quality and providing a solid quantitative foundation for subsequent empirical analyses.

Bias

To address potential sources of bias, several strategies were implemented, with a specific focus on common method bias given the exclusive reliance on self-reported data. Procedural remedies during the research design and data collection phases included anonymous completion, randomization of questionnaire items, and the use of scales with varied response formats to mitigate the influence of potential source bias at the outset. Statistically, Harman single-factor test was planned and used to empirically evaluate the presence of common method bias [].

Study Size

The study size was determined by both feasibility and statistical considerations. Given the constraints of the stratified cluster sampling design, we aimed to recruit a feasible sample from the accessible schools in 5 provinces. The final analytic sample consisted of 855 adolescents. To confirm the adequacy of this achieved sample, a post hoc power analysis was conducted using G*Power 3.1. For a linear multiple regression with 8 predictors, an alpha of 0.05, and a medium effect size (f²=0.15), our sample provided over 95% power, substantially exceeding the minimum requirement of 160 participants, ensuring sufficient power to detect medium and small effect sizes.

Statistical Methods

Data analysis was performed using IBM SPSS 20.0 for tests of common method bias, descriptive statistics, and correlation analysis; IBM Amos 24.0 for CFA and structural equation modeling (SEM); and SPSS PROCESS v4.1 macro for testing mediation and moderation effects. To account for the nonindependence of observations arising from the clustered sampling design, all primary analyses (including SEM and moderated mediation) were conducted using cluster-robust SEs with the classroom as the cluster unit. This method provides corrected SEs and CIs, mitigating the risk of type I error.

A total of 1000 questionnaires were distributed, with 893 returned, resulting in a response rate of 89.3%. The 107 nonreturned questionnaires were primarily attributed to student absences or leave during the survey period. The 893 returned questionnaires were screened for validity and missing data through a 2-stage process. First, to handle missing data, incomplete questionnaires (n=28) with any missing items were excluded. The extent of missing data was minimal, with item-level missingness ranging from 0.2% to 0.5% across all study variables. Little’s test for missing completely at random (MCAR) was nonsignificant (χ²₈₅=98.76, P=.15). Given that the data met the MCAR assumption and the fraction of incomplete cases was very small, the use of complete-case analysis was considered methodologically appropriate for this step [].

Subsequently, a second stage of screening for data quality was performed on the remaining complete cases, leading to the exclusion of an additional 10 questionnaires due to abnormal completion times (n=7) or patterned responses (n=3) []. Following this rigorous process, 855 valid and complete questionnaires were retained for all subsequent analyses, representing 95.7% of returned surveys.

Ethical Considerations

This study was conducted in strict accordance with ethical principles for research involving human participants. The study protocol was reviewed and approved by the Academic Ethics Committee of Capital University of Physical Education and Sports (approval no. 2025A0105). Prior to data collection, written informed consent was obtained from all participating adolescents and their parents or legal guardians. The consent process was conducted through the schools. Specifically, information sheets and consent forms detailing the study’s purpose, procedures, potential risks and benefits, and the right to withdraw at any time without penalty were distributed to parents via students. After parental consent was secured, the same documents were presented to the adolescents in an age-appropriate manner to seek their assent.

All data were collected and processed anonymously, with no personally identifiable information stored, to ensure participant privacy and confidentiality. Data were used solely for the purposes of this academic research. Participants did not receive any monetary or material compensation for their involvement in the study. Finally, this manuscript does not contain any images that could lead to the identification of an individual participant.

Results

Common Method Bias Test

The results of the Harman single-factor test were examined to assess the threat of common method bias. The unrotated exploratory factor analysis revealed 8 factors with eigenvalues greater than 1. The first factor accounted for 31.07% of the total variance, which is significantly below the critical threshold of 40%. This indicates that no single factor explained the majority of the variance, suggesting that common method bias is not a severe issue in this dataset and does not pose a substantial threat to the interpretation of the study’s findings. Furthermore, the CFA demonstrated good psychometric properties for all measurement scales, confirming their construct validity ().

Intraclass Correlation and Clustering Effects

Prior to testing our hypotheses, we quantified the clustering effect by calculating the intraclass correlation coefficient (ICC) for our primary outcome variables. The analysis revealed that 8.5% of the total variance in SMA (ICC=0.085, 95% CI 0.043-0.150) and 6.2% of the variance in DASS (ICC=0.062, 95% CI 0.028-0.125) resided between classrooms, justifying the need to account for the clustered data structure. All subsequent analyses used cluster-robust SEs to ensure valid statistical inference.

Descriptive Statistics and Pearson Correlation Analysis

Descriptive statistics and correlation coefficients for each variable are presented in . The results revealed significant positive correlations between eHL, MHL, and MeHL. Specifically, eHL showed a moderate to strong positive correlation with both MHL and MeHL. MHL and MeHL also demonstrated a significant positive correlation. Furthermore, all 3 health literacy variables were significantly negatively correlated with SMA and DASS, with these relationships being statistically significant. This suggests that individuals with higher health literacy levels tend to experience lower levels of SMA and psychological distress related to DASS. Gender was significantly correlated with some variables, while age showed relatively weaker correlations. Therefore, gender and age were included as control variables in subsequent analyses to enhance the accuracy and robustness of the results.

Mediation Effect Test

To further explore the mechanisms of the associations between eHL, MHL, and SMA among high school students and to examine the potential mediating effect of MeHL, the study constructed an SEM path analysis. The chi-square statistic was significant (χ²₂₂₇₆=6102.464, P<.001); thus, the normed chi-square (χ²/df) was consulted []. The fit indices of the revised model indicated an acceptable to good fit with the data: χ²/df (2.681), root-mean-square error of approximation (0.044), standardized root-mean-square residual (0.048), normed fit index (0.895), incremental fit index (0.931), comparative fit index (0.931), Tucker-Lewis index (0.927), and coefficient of determination (0.85) []. Although the goodness-of-fit index (0.814) and the adjusted goodness-of-fit index (0.797) were slightly below the commonly used threshold (0.90), the model still demonstrated good fit based on the other major fit indices, supporting further analysis of the relationships between the variables [,].

As shown in and , both eHL and MHL were significantly and positively associated with MeHL (β=0.531, 95% CI 0.468-0.594, P<.001; β=0.356, 95% CI 0.297-0.415, P<.001). Both were significantly negatively associated with SMA (β=–0.152, 95% CI –0.242 to –0.062, P=.01; β=–0.261, 95% CI –0.345 to –0.177, P<.001). However, eHL did not have a significant effect on DASS (β=–0.011, 95% CI –0.060 to 0.038; P=.81), while MHL was a significant negative predictor of DASS (β=–0.590, 95% CI –0.647 to –0.533; P<.001). Additionally, MeHL was also a significant negative predictor of SMA and psychological distress (β=–0.150, 95% CI –0.277 to –0.023, P=.02; β=–0.139, 95% CI –0.207 to –0.071, P=.006).

‎

To further examine the mediation effects, we calculated the fully standardized indirect effects and their 95% bias-corrected CIs using bootstrap sampling with 5000 resamples and cluster-robust SEs (). The results indicated that both eHL and MHL exerted significant negative indirect effects on SMA through MeHL, as the 95% CIs for these paths did not contain zero. Specifically, the indirect effect of eHL was β=–0.062, 95% CI –0.125 to –0.010, and that of MHL was β=–0.045, 95% CI –0.090 to –0.008. In practical terms, this suggests that for every one standard deviation increase in eHL, SMA is expected to decrease by 0.062 standard deviations via the mediator. Similarly, the indirect effects on DASS were also significant (eHL: β=–0.037, 95% CI –0.072 to –0.009; MHL: β=–0.027, 95% CI –0.053 to –0.007). The proportion of the total effect mediated (relative effect size) for each path is also presented in .

In summary, the research results are consistent with a model in which MeHL plays a potential mediating role in the associations linking eHL and MHL to adolescent SMA and DASS. Specifically, higher levels of eHL and MHL were associated with higher MeHL, which in turn was linked to lower levels of SMA and psychological distress related to DASS. The study suggests a potential synergistic pathway of the “eHealth–mental health” dual-track mechanism that may be relevant for interventions targeting SMA and psychological distress, offering significant theoretical value and practical implications.

Moderation Effect Analysis

To explore the interaction between eHL and MHL in predicting MeHL, the study used a stepwise regression analysis with cluster-robust SEs to construct a 3-stage progressive model (). Model 1 serves as the baseline model, controlling for gender and age. The results show that demographic variables explain very little of the variance in MeHL (R²=0.008), though the overall model was not statistically significant (F=3.344, P<.05). Model 2 introduces eHL and MHL as predictors. Both have a significant positive effect on MeHL (eHL: β=0.535, 95% CI 0.472-0.598, P<.01; MHL: β=0.288, 95% CI 0.229-0.347, P<.01). The model’s explanatory power significantly increases (R²=0.544), with a ΔR²=0.536 compared to model 1, F=253.232, P<.001, indicating that these 2 forms of health literacy are important predictors of MeHL. Model 3 adds the interaction term of eHL and MHL (centered). The results show that the interaction term’s regression coefficient is significant (β=0.067, 95% CI 0.010-0.124; P<.05), and the model’s explanatory power slightly increases (R²=0.547, ΔR²=0.003, F=205.285, P<.001), indicating a synergistic effect between the two, enhancing their ability to predict MeHL.

To further clarify the nature of the interaction, a simple slope analysis was conducted to probe the association between eHL and MeHL at low (mean – 1 SD), medium (mean), and high (mean + 1 SD) levels of MHL (). The analysis revealed a significant positive association between eHL and MeHL at low (b=1.040, 95% CI 0.881-1.199), medium (b=1.154, 95% CI 1.033-1.275), and high (b=1.268, 95% CI 1.131-1.405) levels of MHL. As visually represented in , which plots the predicted MeHL scores and their CIs, these results demonstrate a clear synergistic pattern: the positive relationship between eHL and MeHL grows progressively stronger as the level of MHL increases.

‎

Moderated Mediation Effect Test

The study used the PROCESS (model 7) model proposed by Hayes in 2022 to test the mediation mechanism of MeHL in the relationship between eHL and SMA, as well as the moderating effect of MHL on this mediation path. In the analysis, all continuous predictor variables (ie, eHL and MHL) were mean-centered to reduce the impact of multicollinearity on the estimation of interaction terms. The bootstrap method was used for 5000 resamples to generate 95% bias-corrected CIs to assess the significance and stability of the moderated mediation effect. The Johnson-Neyman technique was used to precisely characterize the moderation pattern across the full range of MHL levels.

The moderated mediation analysis results (), estimated with cluster-robust SEs, demonstrate a consistent strengthening of indirect effects with increasing MHL levels. For SMA, the standardized indirect effect intensified from β=–0.084 at the 16th percentile to β=–0.099 at the 84th percentile of MHL, with a statistically significant pairwise contrast between high and low levels (Δ=–0.016, 95% CI –0.038 to –0.006). A more pronounced pattern was observed for DASS, where the indirect effect increased from β=–0.252 to β=–0.299 across the same percentile range, supported by a significant pairwise contrast (Δ=–0.047, 95% CI –0.102 to –0.008). The Johnson-Neyman analysis further confirmed that these mediation pathways remain statistically significant throughout the complete observed range of MHL, with no significant transition points identified. These findings indicate that higher levels of MHL significantly enhance the protective indirect effect of eHL on both behavioral and psychological outcomes through MeHL.

Discussion

Principal Findings

This study proposed and tested a multidimensional health literacy model to elucidate the synergistic pathways through which eHL and MHL influence adolescent SMA and DASS, with MeHL as a mediator and MHL as a moderator. Our findings, supported by robust bootstrap CIs, provide precise estimates of these complex relationships. In support of our hypotheses, the findings revealed that (1) H1 was partially supported, as both eHL and MHL were significant negative predictors of SMA, with CIs indicating stable negative associations, but only MHL was a significant negative predictor of DASS, while the nonsignificant path from eHL to DASS suggests a negligible direct effect; (2) H2 was fully supported, with MeHL serving as a significant mediator in the relationships from both eHL and MHL to SMA and DASS, with all indirect effect CIs excluding zero; and (3) H3 was also supported, as MHL positively moderated the effect of eHL on MeHL, indicating a statistically significant though modest synergistic interaction that strengthened the proposed moderated mediation pathways to the outcomes.

Delving into the distinct associations posited in H1, our findings clarify the unique roles of different health literacies. These differential pathways align with yet extend previous research that has often examined these literacies in isolation. MHL demonstrated a robust negative association with DASS, suggesting it serves as a primary psychological resource for emotional regulation [], effectively breaking the “negative emotions–compensatory use” cycle [,]. This is consistent with population-based surveys that have identified MHL as a key correlate of psychological well-being in adolescents []. Conversely, eHL showed a more targeted association with reducing behavioral risks (SMA), equipping adolescents to navigate algorithmic “information cocoons” and use practical strategies like time management [,]. This finding resonates with large-scale studies linking digital competency to lower risks of problematic online behaviors, highlighting the role of eHL in mitigating specific behavioral risks rather than broad emotional distress []. Together, these dual pathways illustrate how distinct health literacies serve as protective resources. This mechanism aligns with the COR theory, as both literacies help adolescents conserve mental resources and mitigate digital threats []. Therefore, enhancing multidimensional health literacy constructs a dual barrier of “cognitive defense–behavioral regulation,” providing concrete intervention entry points.

The mediating role of MeHL (H2) represents a crucial synthesis of digital capability and psychological cognition. Our bootstrap analysis provides precise estimates of this integration, showing significant indirect effects from both eHL and MHL through MeHL. This result is consistent with the “triadic reciprocal model” in SCT, where personal factors (MHL), behaviors, and the digital environment interact continuously []. eHL provides the tool for information access, while MHL provides the cognitive framework for interpretation; their confluence fosters MeHL. This integration is critical, as individuals must not only find digital health information but also critically evaluate and apply it. The ability to identify and resist health misinformation is a key component of advanced health literacy, which our MeHL construct appears to capture []. Furthermore, general health literacy has been associated with more sophisticated information processing strategies, which may underpin the effective integration of digital and mental health knowledge reflected in MeHL []. Notably, the stronger mediation effect for eHL suggests the digital capability pathway is particularly dependent on this integrative process. MeHL thus plays a key role in the “knowledge internalization–behavioral transformation” process, offering a layered intervention path.

Most importantly, the supported moderated mediation hypothesis (H3) reveals a dynamic synergy at the heart of our model. The positive moderation effect indicates that higher MHL levels strengthen the effect of eHL on MeHL. This finding offers strong support for COR theory’s principle of resource caravans [], wherein existing resources (MHL) facilitate the acquisition and transformation of new resources (eHL). This boosting effect makes sense when we consider teenage brain development. Teenagers’ brains are especially tuned to seek rewards and are highly motivated to learn new things that help them adapt []. In this context, MHL may provide the necessary motivational and cognitive framework—the “why”—making the technical skills of eHL—the “how”—more valuable and worth applying to mental health contexts, thereby enhancing MeHL. This positions MHL as a “cognitive catalyst” that unlocks digital skills’ potential. Individuals with high MHL show greater cognitive processing depth when dealing with digital health information, while those with low MHL struggle to achieve behavioral transformation despite technical skills [].

In conclusion, this study challenges the siloed perspective of previous research by proposing and validating a comprehensive model that elucidates the dynamic interplay between different health literacies. Unlike earlier work that primarily examined eHL and MHL in isolation, our findings demonstrate that their synergistic association—where MHL catalytically enhances the utility of eHL—is crucial for understanding adolescent outcomes in the digital age. This integrated perspective is supported by research indicating that combined literacy interventions can be more effective, and our model provides a specific mechanistic account of this synergy []. This validated dual-path moderated mediation model provides a significant theoretical advancement by moving beyond direct effects to reveal a more complex relational network. Methodologically, our robust approach accounting for clustered data strengthens the ecological validity of these observed associations. These insights carry direct practical implications, providing an evidence-based blueprint for integrated health promotion that concurrently targets psychological cognition and digital competencies. Fostering this synergy represents a promising and efficient strategy for safeguarding adolescent well-being and building a healthier digital future.

Limitations

Notwithstanding its contributions, this study has several limitations that should be considered. The cross-sectional design precludes causal inference, leaving open the possibility of reverse causality, such as the potential for psychological distress to diminish health literacy acquisition. While common method bias was statistically assessed and fell below critical thresholds, the use of self-reported data may still introduce social desirability bias and inflate relationship estimates. Furthermore, the generalizability of the findings may be constrained by the regional sampling frame within China, and the omission of potential confounders such as socioeconomic status, academic pressure, and precise social media usage metrics means that the reported associations, though robust, may partially reflect the influence of these unmeasured variables. Future longitudinal studies incorporating objective behavioral data and a broader set of covariates in diverse cultural contexts are warranted to confirm the causal and generalizable nature of these synergistic pathways.

Conclusions

This cross-sectional study validates a new model that overcomes the limitations of previous research, which typically examined health literacy dimensions in isolation. It reveals the key role of eHL and MHL in reducing SMA and alleviating DASS among adolescents. The study shows that while both eHL and MHL contribute to reducing SMA, only MHL directly alleviates DASS. Additionally, MeHL plays a critical mediating role in this process. The innovation of this study lies in demonstrating that MHL is not merely an independent protective factor but an important moderator that enhances the effect of eHL in promoting MeHL. MHL provides the cognitive foundation for the effective application of digital skills, enabling adolescents to better use digital tools to manage their mental health, thereby maximizing the protective potential of digital skills. Therefore, public health strategies should focus on integrated interventions that simultaneously enhance both psychological resilience and digital competence, which is crucial for safeguarding adolescent well-being.

Introduction

Adapted physical activity (APA) is an evidence-based, multidisciplinary rehabilitation program currently considered the gold standard treatment for obesity or type 2 diabetes (T2D) because of its multiple benefits such as weight and fat loss and improvements in blood pressure, cardiorespiratory fitness, insulin sensitivity, appetite control, and quality of life []. The economic impact of such programs could be substantial as, in France, the estimated health care savings per individual who becomes sustainably active amount to €840 (US $967.86) for individuals aged 20 years to 39 years and €23,275 (US $26,817.70) for individuals aged 40 years to 74 years [].

However, despite these well-documented benefits, APA programs often struggle with low adherence and observance rates [], with reported dropout rates reaching approximately 50% [,]. Moreover, only a low percentage of patients maintain their physical activity (PA) at the end of these programs [-]. For instance, a systematic review published in 2017 revealed that only 33% of participants adhered to prescribed exercise programs following the completion of supervised training []. Logistical and financial barriers further limit the impact of these programs, including factors such as inconvenient timing, transportation difficulties, program cost, and location [,]. Additional challenges such as lack of self-discipline or motivation, pain or physical discomfort, and lack of time are generally mentioned by patients with obesity [].

Given these challenges, there is increasing interest in leveraging digital health technologies for obesity treatment. Digital health interventions hold promise to better change behavior in real-life contexts and have the potential to extend the reach of evidence-based behavioral interventions at lower cost while reducing patient burden []. Multiple interventional perspectives emerge with the democratization of digital tools. Among these innovations, gamification stands out; it is defined as the use of game mechanisms such as points, levels, leaderboards, badges, challenges, or customization in non-game contexts to foster behavior change, engagement, and motivation and to solicit participation. Recent meta-analyses suggest that gamified interventions are effective at promoting long-term changes in PA among various participants, including clinical populations [-]. Notably, compared with digital interventions that do not incorporate game mechanics, gamified interventions produce larger effects on PA (g=0.23) [] and improvements in daily step count (+489 steps/day), BMI (−0.28 kg/m²), body weight (−0.70 kg), body fat percentage (−1.92%), and waist circumference (−1.16 cm) []. These findings suggest that gamified interventions not only promote behavioral change but also enhance the overall effectiveness of traditional digital health programs. In addition, telehealth solutions are emerging as a safe alternative for proposing remote APA sessions with good acceptability among patients and positive clinical results [-].

Nevertheless, despite the promise of digital health interventions, the evidence supporting their use in obesity treatment remains limited. Although digital interventions have been extensively tested in comparison with passive or active control groups, no rigorous trial has yet demonstrated the superiority of digital PA interventions over existing ones (eg, supervised and APA programs). In addition, evidence regarding long-term effects and cost-effectiveness remains inconclusive [-]. In this context, the digital health domain needs more rigorous evaluations, including follow-up assessments, comparisons with usual care, and economic evaluations.

In addition, although the gold standard evaluation method, namely the randomized controlled trial (RCT), can be helpful for evaluating the efficacy and effectiveness of digital health tools [,], evaluation processes must extend beyond the average response in order to consider the idiosyncratic nature of PA (ie, the fact that people’s activity and responses to an intervention can largely differ from one person to another) []. This emphasizes the need to adopt an idiographic approach (ie, individual statistical modeling) and report both between-group differences in benefit and within-person evolutions []. Finally, as the PA behavior change process is dynamic, it requires high-resolution measurements in order to capture its relative variability or stability over time [,]. Since intervention effects can be nonlinear and vary across time scales, the traditional measurement bursts conducted in PA trials (ie, 7-day measurement before and after the intervention) are not adapted to properly assess the impact of a behavioral change intervention on PA []. To answer these 2 challenges, integrating intensive longitudinal measures with N-of-1 methods within an RCT may prove particularly valuable [].

Herein, we present the results of the Digital Intervention Promoting Physical Activity among Obese people (DIPPAO) study, an RCT designed to compare a digital intervention integrating gamification and telecoaching (the Kiplin program) with usual care. The study tests the hypothesis that these 2 interventional features can augment PA levels during and at the end of the program – ultimately leading to better clinical outcomes in the long haul. To address limitations observed in previous studies, this trial included (1) a 6-month follow-up to assess the sustainability of the intervention effect after the end of the intervention, (2) a cost-utility analysis to assess the economic impact of both programs, (3) an intensive longitudinal assessment of daily steps throughout the entire study to capture daily PA behavior more precisely, and (4) an analysis of both between-group differences and within-person evolutions using an idiographic approach.

Methods

Overview

The DIPPAO study is a 2-arm parallel prospective randomized controlled trial conducted in France. Full details of the study methods have been reported previously []. This trial aimed to compare the effectiveness, cost-effectiveness, and psychological mechanisms of the Kiplin program in comparison with face-to-face, supervised PA sessions among patients treated for obesity or T2D. The authors are solely responsible for the design and conduct of this study; all analyses, drafting, and editing of the paper; and its final contents. In this paper, we report the primary and secondary clinical outcomes along with the cost-utility analysis; the psychological mechanisms results are reported separately. There were no significant deviations from the prespecified protocol during the trial. The reporting follows the Consolidated Standards of Reporting Trials (CONSORT) guidelines [] and the Template for Intervention Description and Replication checklist [].

Ethical Considerations

Ethics Approval

The protocol of this trial was reviewed and approved by the National Human Protection Committee (CPP Ile de France XI, number 21 004‐65219). All eligible individuals were informed about the study’s objectives, procedures, and ethical considerations, including assurances of confidentiality, and provided written informed consent prior to participation.

Compensation

Participants did not receive monetary compensation; however, the wearable device distributed at the beginning of the study was offered to them upon its completion.

Privacy and Confidentiality

Our trial was conducted in accordance with established standards for privacy and confidentiality. All investigators with direct data access took appropriate measures to safeguard the confidentiality of information related to the medical products, the trial, and the participants, particularly their identity and outcomes. No identifiable patient information is included in the manuscript, the supplementary materials, or the project’s Open Science Framework page.

Study Population, Context, and Procedure

Context and Settings

The study was conducted within the department of sports medicine of the University Hospital of Clermont-Ferrand, France, located in an urban area with approximately 500,000 inhabitants. The hospital is situated within the city and is easily accessible by car and public transportation. The unit admits around 450 patients with various chronic diseases annually and is equipped with a dedicated clinical research office, electronic health record systems, and specialized facilities for supervised APA sessions. Participants were referred to the program by their general practitioners. The intervention programs were provided free of charge to participants; however, transportation to the hospital was at their own expense. Patients’ availability for participation was strongly influenced by their occupational commitments. In-person sessions were offered at multiple times throughout the week to maximize feasibility for the patient.

Participants

Patients were screened between June 2021 and October 2022. Eligible patients were 18 years to 65 years old who were treated for obesity (BMI ≥30 kg/m² and <45 kg/m²) or overweight or obesity and T2D within the department of sports medicine. Patients were required to own either an Android-based phone or Apple iPhone with a study-supported operating system. Full inclusion and exclusion criteria are available in Table S1 in . Participants attended 5 visits throughout the study: a selection visit, an inclusion visit, and 3 experimental visits.

Selection and Inclusion Visits

During the selection visit, one of the investigating physicians checked the patients’ eligibility criteria and signed the informed consent form. Only after providing informed consent did participants proceed to the inclusion visit, where they were given a wearable device (Garmin Vivofit 3; Garmin International) and an accelerometer (Actigraph GT3x; ActiGraph LLC) for baseline assessment of PA over a 7-day period. Randomization was performed by the associate biostatistician using a permuted block design with variable block sizes and a 1:1 allocation ratio. The randomization list was transmitted using sequentially numbered, opaque, sealed envelopes to the data collectors.

Experimental Visits

The T0 experimental visit marked the baseline assessment conducted prior to the start of the intervention. At the conclusion of the 3-month program, participants completed the T1 experimental visit. To evaluate the long-term effects of the intervention, the T2 experimental visit was conducted 6 months postprogram completion. Research assistants collecting data were blinded to treatment allocation. However, double blinding was not feasible in this context because the intervention’s nature made it impossible to conceal allocation from the participants.

Intervention Overview

Intervention Group

Details on the intervention content have been reported previously [,]. The Kiplin intervention is a theory-driven digital APA intervention grounded in the social identity approach and self-determination theory. It comprises 4 key components: (1) 22 APA sessions (2 face-to-face sessions at the program’s start, followed by 20 remote videoconference sessions); (2) 3 PA collective games accessible through an Android or iOS app; (3) chat feature allowing communication with other participants, teammates, and health professionals; and (4) an activity monitoring tool enabling participants to track their PA in real time (). The intervention integrates a total of 16 behavior change techniques and primarily aims to promote changes in daily PA behavior.

Participants completed 3 sessions per week during the first 2 weeks (1 face-to-face and 2 telecoaching sessions), 2 telecoaching sessions per week during the following 6 weeks, and 1 telecoaching session per week during the third month, for a total of 22 sessions. The 2 initial face-to-face sessions were conducted at the University Hospital under the same conditions as the control group (see the following paragraphs) and were designed to ensure that participants correctly adopted the required movements. The telecoaching sessions consisted of 60-minute, group-based live APA classes delivered remotely by a professional APA coach to small groups of 5 to 7 participants. These participants played the Kiplin games together then met for the group sessions, fostering both social connection and engagement. Each week, several session time slots were available, allowing participants to register according to their preferences and availability. Sessions incorporated interactive elements such as quizzes, riddles, and practical tips on PA, in addition to endurance, muscle strengthening, and stretching exercises, as well as therapeutic education.

Participants engaged in 3 distinct games, with no option for selection. In all games, participants’ daily step counts were converted into points, enabling progression within the game environment. The Kiplin app retrieved participants’ daily step counts via integration with the application programming interface of the tracking app (Garmin Health in this study). In The Adventure, the objective was to collectively achieve step goals to progress toward a final destination; players tracked their progress on a digital world map with checkpoints representing distances between cities (). In The Mission, participants completed PA and collective challenges to unlock clues and solve missions (). Finally, in The Board Game, participants acted as forest rangers tasked with extinguishing fires. Meeting step goals enabled progress on the board and advancement to higher levels, with the ultimate goal of extinguishing all fires and saving forest residents ().

These games incorporated multiple gamification mechanisms such as points, trophies, leaderboards, chat features, challenges, and narratives (see [] for an overview of Kiplin’s gamification strategies following the taxonomy proposed by Schmidt-Kraepelin et al []). The telecoaching APA coaches also participated in the games alongside the participants and were available to answer game-related questions at the end of telecoaching sessions.

Control Group

Participants allocated to the control condition received the traditional 3-month face-to-face, supervised APA program at the University Hospital of Clermont-Ferrand, with 3 sessions a week on nonconsecutive days for a total of 36 sessions. The sessions were conducted individually, with no interaction between participants. Each session included a warm-up and 50 minutes of endurance and muscle-strengthening exercises, followed by stretching, all under the supervision of an APA coach in a dedicated facility. Aerobic and resistance exercises were organized in a circuit comprising 6 exercise stations (3 aerobic and 3 resistance). Aerobic exercises were performed at 50% of VO₂max during the first week, with intensity progressively increased by 10% every 2 weeks to reach at least 80% of VO₂max during the final 9 weeks. Resistance exercises were performed at 50% of 1-repetition maximum (1RM) in the first week, with the load similarly increased by 10% every 2 weeks to reach and maintain 80% of 1RM during the final 5 weeks.

Study Outcomes

Full details of the outcome measures, including their reliability and validity, have been reported previously []. At baseline, participants provided demographic data, including date of birth, sex, and highest level of education completed. The primary outcome was the change in daily step count, assessed at high resolution using the Garmin Vivofit, a wearable activity tracker with validated accuracy under various walking conditions [], from baseline to 1-week postintervention. Additionally, the change in daily steps in the follow-up period from baseline was also evaluated.

The 3 experimental visits included the following clinical outcomes. BMI was calculated as body mass (kg) divided by height squared (m²). Body composition, including fat and lean mass, was assessed using bioelectrical impedance analysis with the multifrequency segmented body composition analyzer (Tanita MC780; Tanita). Moderate-to-vigorous PA (MVPA), light PA (LPA), and sedentary time were measured over 7 consecutive days using a triaxial accelerometer (ActiGraph GT3x). Cardiorespiratory fitness was assessed via the 6-minute walk test (6MWT). Muscular strength of the upper limbs was assessed via handgrip strength using a dynamometer (Takei Grip-D; Takei) with a series of 3 handgrip test measurements for right and left hands in the seated position. Muscular strength of lower limbs was assessed through maximum knee extension torque using an isokinetic dynamometer at speeds of 30 °/s, 60 °/s, and 120 °/s. Quality of life was measured via the EQ-5D-5L questionnaire [], which evaluates the following 5 dimensions: mobility, autonomy of the person, current activity, pain and discomfort, and anxiety and depression.

For both groups, the number of APA sessions attended was assessed. For the Kiplin group, app engagement was collected and comprised participation in the games, app usage frequency, and the number of messages sent.

The health economic evaluation was conducted through a cost-utility analysis incorporating the (1) identification and valuation of costs and (2) measurement of utility using the EQ-5D questionnaire. The analysis was performed from the health insurance perspective considering only direct medical costs. The time horizon for this evaluation extended from baseline (T0) to the follow-up assessment at T2.

Statistical Analysis

The trial was designed to demonstrate a difference equivalent of an effect size of d=0.6 on our primary outcome (daily steps) for 80% power and a 2-sided type I error at 0.05. In idiographic designs, statistical experts [] recommend a minimum of 835 observations for adequate power (.80) at level 1. With 2058 (98 days × 21 participants) observations, we had adequate power to detect within-person changes in the Kiplin condition [].

All analyses followed a modified intention-to-treat principle, incorporating participants with complete data for the primary or secondary outcomes at the initial 2 experimental visits. This pragmatic analytical approach aimed to enhance internal validity by assessing treatment effects among participants who received the intervention while acknowledging that the necessary exclusions could attenuate some benefits of randomization. Analyses were performed using R (R Foundation for Statistical Computing). Baseline variables are reported as numbers and percentages for categorical variables and means with SDs for continuous variables. According to the CONSORT 2010 statement [], group differences in baseline variables were not compared using significance testing. Two authors directly accessed and verified the underlying data reported in the manuscript, including one who had no affiliation with the Kiplin company.

Changes in daily step count were assessed between baseline and intervention periods as well as between baseline and follow-up using linear mixed-effects models. These models accounted for the nested data structure and included fixed effects for group, time, and group × time interaction. Because plots of the residuals and estimated trajectories suggested substantial nonlinearity in step count change over time, we compared polynomial extensions of the time effect (quadratic, cubic, and quartic) by minimum Akaike information criterion (AIC) and Bayesian information criterion (BIC). The quadratic model provided the best fit and is presented in the Results section. Random intercepts for participants and random linear slopes for repeated measures at the participant level were included. As an additional analysis, we treated measurement occasion as a categorical “period” factor (baseline, intervention, follow-up) rather than as a continuous measure of days. In this model, we included fixed effects for group, period, and the group × period interaction with the same random effects structure.

Adjusted models were further fitted and incorporated fixed effects for group, period, group × period interaction, age, baseline PA levels, BMI, season, and number of completed APA sessions – factors previously identified as significant predictors of intervention effects []. Nonwear days, defined as days with fewer than 1000 steps, were treated as missing data. All models were performed using the lmerTest package [], and contrast analyses were conducted using the emmeans package []. Diamond comparison plots were drawn with the R package ufs [].

In a complementary analysis of the primary outcome, an idiographic approach was used to analyze daily step counts for each participant separately using generalized additive models (GAMs) []. GAMs are an extension of generalized linear mixed models that allow the estimation of smoothly varying trends where the relationship between the covariates and the response is modeled using smooth functions []. GAMs are particularly well-suited to the modeling of time series data with 1 level of measurement (ie, repeated measurements nested within 1 individual), as they can accommodate the inclusion of autocorrelated error terms []. Nonlinearity was assessed via the effective degrees of freedom (edf) of smoothing terms, with edf ≥3 indicating nonlinearity []. GAMs were computed using the mgcv package [], and the visreg package [] was used for model visualization.

Continuous secondary outcomes were analyzed using mixed-effect models (including the group, time, and group × time interaction terms as fixed effects) to assess changes across 9 months (ie, considering the 3 measurement points). As statistical power was computed for the primary outcome, secondary outcome analyses were classified as exploratory [].

Results

Participant Characteristics

Between June 22, 2021, and October 6, 2022, 57 patients were screened for eligibility, with 50 meeting the inclusion criteria and subsequently randomized to either the Kiplin group (n=25) or usual care group (n=25; ). Participants who completed at least 2 assessments (modified intention-to-treat sample) had a mean age of 47.7 (SD 13.3) years and a mean BMI of 40.0 (SD 7.14) kg/m², and 31 were women (31/42, 74%). A total of 13 participants had T2D (13/42, 31%). At baseline, the average daily step count was 6584 (SD 3787) steps per day. No substantial differences in baseline characteristics were observed between groups (). The descriptive statistics of all the randomized participants are available in Table S2 in .

‎

Compliance and Engagement Metrics

Among the 50 randomized participants, 8 withdrew or were lost before program completion (3 in the intervention group and 5 in the control group), and 17 were lost during the follow-up period (). Of a possible 274 days, participants logged steps for a median of 158 (IQR 162) days throughout the study, with 80% (34/42) of participants wearing their activity monitor for at least 75% of days. In the Kiplin condition, participants attended an average of 14.68 (SD 6.14) of 22 possible APA sessions, yielding a completion ratio of 0.67 (SD 0.28). They engaged in an average of 2.6 (SD 0.71) of a possible 3 PA games, corresponding to a mean of 36.6 (SD 10.02) days in-game; sent an average of 14.2 (SD 15.93) messages; and logged into the app on 88.12 (SD 43.63) times on average. In contrast, patients receiving usual care attended an average of 30.27 (SD 5.88) of 36 APA sessions, with a completion ratio of 0.84 (SD 0.16).

Primary Outcome

Between-Group Differences

During the 3-month intervention period, participants in the Kiplin condition increased their daily steps by 1092.06 (95% CI 533-1651.5) steps per day on average (+10.9%; P<.001; d=0.38) compared with the baseline period and by 834 (95% CI 114-1554.7) steps per day on average during the follow-up period (+8.3%; P=.01; d=0.29) compared with baseline. In contrast, the daily step count of the usual care condition remained stable and showed no significant changes in daily steps during the intervention period (+0.1%; P≥.99) or follow-up period (+0.9%; P=.10) relative to baseline. depicts the unadjusted daily step count evolution by phase and condition.

Results of our quadratic mixed-effects model predicting daily step count from time (days since baseline) indicated that the time trajectory of daily steps differed significantly between the Kiplin and usual care conditions (). The usual care group exhibited a nonsignificant linear decline of 20.06 steps per day (P=.07), followed by a small but significant upward curvature (ß=0.126, P=.03). In contrast, the Kiplin group’s linear slope was 41.98 steps per day steeper than than that of the control (P=.004), and its quadratic term was 0.229 steps/day² lower (P=.005).

‎

‎

In our mixed-effects model predicting daily step count from period (baseline, intervention, follow-up), the contrast analyses revealed that participants of the Kiplin group had a significantly greater increase in mean daily steps between baseline and the intervention period compared with the usual care group, with an estimated difference of 1085 (95% CI 493-1676) steps (P<.001) relative to baseline. Similarly, during the follow-up period, the daily step count change from baseline for participants in the Kiplin group remained significantly greater, at 1775 (95% CI 906-2643) steps (P<.001).

Results of the adjusted models follow the same trends and are available in Tables S3 and S4 in .

Within-Person Evolutions

The results of the GAMs analyzing the evolution of daily step counts over time for each participant in the Kiplin group are summarized in Table S5 in . The edf for the smoothing terms ranged from 1.0 to 8.10, with 9 participants exhibiting edf>3, indicative of nonlinear patterns. Among the participants, 8 showed significant changes in daily step counts over time (all P<.05; see Table S5 in for the P values), with 7 showing improvements and 1 experiencing a decline. Visualizations of the daily step trajectories for these 8 participants are presented in , with the complete set of plots available in Figure S1 in .

‎

Secondary Outcomes

presents the results of the secondary outcome measures. Within-group comparisons showed significant improvements from baseline to 9 months in lower limb muscle strength (ß=6.68, 95% CI 1.40 to 11.96; P=.01) and a significant decrease in BMI at 3 months (ß=−0.51, 95% CI −0.98 to −0.04; P=.03) for the Kiplin group. In contrast, the control group showed significant improvements in quality of life at 3 months (ß=0.08, 95% CI 0.01 to 0.15; P=.02) and in 6MWT and lower limb muscle strength at both 3 months (ß=38.94, 95% CI 18.96 to 58.93; P<.001, and ß=9.27, 95% CI 4.15 to 14.40; P=.001, respectively) and 9 months (ß=25.41, 95% CI 2.04 to 48.78; P=.03, and ß=7.59, 95% CI 1.80 to 13.39; P=.01, respectively). Results of the mixed-effect models revealed no significant group × time interactions for secondary outcomes, except for a significant interaction in MVPA change from baseline favoring the Kiplin group (ß=5.69, 95% CI 0.61 to 10.77; P=.03). provides a visual comparison of changes between conditions with effect size estimates. Graphs depicting the evolution of secondary outcomes are available in Figure S2 in .

‎

Economic Evaluation

The cost per patient in the usual care group was €491.92 (US $566.80) compared with €568.74 (US $655.31) in the Kiplin group, resulting in a cost differential (surplus of Kiplin compared with the control group) of €76.82 (US $88.51). Quality-adjusted life years (QALYs) did not significantly differ between groups and time points. The QALYs calculated from the EQ-5D questionnaire were, on average, rather good at baseline, with averages of 0.62 at T0 and 0.67 at T1 and T2.

Discussion

Principal Findings

The results of this RCT revealed that a group-based gamified digital intervention significantly increased daily PA compared with a traditional supervised, face-to-face APA program. Interestingly, this significant difference was observed both during the 3-month program (+1085 daily steps) and during follow-up periods (ie, 6 months postintervention: +1775 daily steps). These results remained consistent irrespective of adjustments for potential confounding factors, reinforcing the robustness of the findings.

In terms of trajectory, participants in the usual care group followed a U-shaped trajectory: their daily steps declined during the intervention and then returned to baseline by the program’s end, which is consistent with our hypothesis and the compensation mechanisms often observed between supervised and leisure-time PA in APA programs []. In contrast, Kiplin participants exhibited an inverted U-shaped pattern, with a sharp early surge in steps that gradually decelerated over time, suggesting the intervention was most potent at the outset.

Comparison With Existing Literature

Our findings are consistent with previous research showing that gamified digital interventions can lead to meaningful real-world improvements in daily step counts, with effects lasting several months after the intervention [,,]. We observed that the intervention’s impact was strongest at the beginning of the program, aligning with evidence that shorter gamified interventions tend to produce the largest gains [,]. Although the follow-up period in our study was longer than the average duration reported in previous research, the effect size 6 months postintervention (834 additional daily steps on average) observed in this study was greater than that of earlier trials [].

To the best of our knowledge, no gamified digital program has previously been compared with usual care in the management of obesity and T2D. However, a prior meta-analysis comparing digital health interventions with minimal intervention or usual care in the context of cardiac rehabilitation found no significant effect on objectively measured PA []. In contrast, our study demonstrated that the digital intervention promoted greater increases in both daily step counts and MVPA compared with usual care. This effectiveness for MVPA is also novel, as previous meta-analyses did not report significant effects of gamified interventions on such outcomes [].

Clinical and Methodological Implications

These results are particularly promising for several reasons. Where current traditional face-to-face, supervised APA programs often face challenges for driving sustained increases in PA [], the Kiplin program addresses this gap effectively. Notably, the effect size observed during the intervention was maintained throughout the 6-month follow-up, underscoring the sustainability of the PA behavior change. The benefits of improving daily PA are now well-recognized []. A recent meta-analysis revealed that taking more steps per day was associated with a progressively lower risk of all-cause mortality, regardless of age, health status, or intensity []. This suggests that the observed behavioral changes in the Kiplin group could have significant long-term health implications.

Nevertheless, in parallel to these results, idiographic analyses revealed substantial variability in individual responses within the Kiplin condition. More especially, GAM models showed (1) significant between- and within-person variability during the intervention, with some participants displaying highly nonlinear patterns while others showed linear trends, and (2) divergent responses to the intervention, with several participants experiencing no significant changes across time. These findings introduce important nuances to our results, indicating that digital interventions may not be suitable for every individual. This underscores the need for effective screening methods to identify patients who are most likely to benefit. Indeed, we can assume that factors such as the stage of behavior change, the acceptability of the technologies [], or some physiological characteristics [] could play a critical role in determining whether a digital or in-person program would be more appropriate.

In addition, no statistical differences were observed in the secondary outcomes evaluated in this study, except for the MVPA change. Although these results should be interpreted with caution, as they are exploratory and stem from secondary analyses potentially without appropriate statistical power, they illustrate 2 different approaches to chronic disease management: behavior change facilitated through the Kiplin program versus functional and physical fitness improvement achieved through the traditional program.

According to the cost-utility analysis, the cost surplus of approximately €76 per patient generated by Kiplin is not significant for an unimpaired quality of life. The anticipated gains in QALYS could not be demonstrated, probably due to the limited sample size at T2 and the high baseline quality of life score reported by patients at T0. This suggests that the generic EQ-5D measurement tool, while useful for QALY transposability, may lack the specificity needed to detect quality of life changes in this population of patients. To provide a conclusive evaluation of Kiplin’s cost-effectiveness over usual care, future research should conduct more specific medical-economic analyses, incorporating quality of life metrics adapted to the patient population. Although digital interventions are often expected to be more cost-effective than in-person programs, the intervention in this study proved to be more expensive. The higher costs observed in the Kiplin group primarily reflect the fixed expenses associated with implementing a hybrid digital intervention, including software licensing, human resources, and equipment. Given the relatively small number of participants in our study, these fixed costs could not be offset through economies of scale. In a larger-scale implementation, the average cost per participant would likely decrease substantially, making the digital component more cost-efficient in practice. Nevertheless, although unexpected, our findings are consistent with previous studies that have not demonstrated clear evidence of cost-effectiveness [,] .

From a methodological perspective, the variability observed with the GAMs reflects a broader limitation of traditional RCTs, which typically focus on group-level differences, potentially overlooking meaningful interindividual variability. By relying solely on aggregate measures, RCTs may fail to capture distinct response patterns, limiting insights into underlying mechanisms and moderating factors that influence intervention effectiveness. In this line, idiographic approaches appear to be a valuable complement to traditional clinical experimental designs, and this study highlights the importance of combining both between- and within-person approaches for evaluating digital interventions, as well as the advantage of high-resolution behavior measurement.

Limitations and Perspectives

However, the results should be interpreted in light of several limitations. First, although using wearables allowed for continuous daily step monitoring in real-world conditions, variations in wear time between groups at the day level cannot be excluded, as the Kiplin group may have been more incentivized to wear the device through gamification. Qualitative interviews (not reported in this paper) with the patients suggested that participants in both groups wore the devices consistently (since recharging was not required), but we lack objective data to confirm this. In addition, these activity monitors are not able to assess several forms of PA, such as cycling or ergometer exercise. Future studies using wearables with continuous heart rate tracking could better assess and control for wear time, and devices such as Motus or SENS may allow for large-scale, 24-hour movement behavior assessment [], including cycling []. Second, results from secondary outcomes should be viewed with caution due to potentially insufficient power to detect these effects as the power test was computed for the primary outcome according to a large effect size. Third, we observed baseline differences in daily steps between the 2 conditions, which were driven by a few participants with unusually high activity for people with obesity or T2D, as our sample reflected real-world hospital admissions without activity-based selection criteria. Future studies could screen for initial PA to minimize such variability. Last, digital interventions consist of a complex interplay of interconnected components, making it challenging to isolate the specific influence of individual elements when the intervention is evaluated as a whole. To address this challenge, innovative research frameworks like the Multiphase Optimization Strategy (MOST) [] or hybrid designs [] could offer a promising avenue in future research for systematically isolating the intervention components and optimizing digital health interventions.

Conclusion

This study confirms the potential of digital health interventions to promote sustained changes in PA compared with usual care in participants with obesity and T2D. Although this behavior change has not led to superior clinical outcomes compared with usual care in this study, its continued persistence beyond the 6-month postintervention period could lead to more pronounced long-term benefits. To validate this hypothesis, larger clinical trials with extended follow-up durations are necessary. This study also had the originality of incorporating both between-group and within-person analyses of daily step counts. The findings indicate that, although the digital intervention effectively increased daily steps on average compared with usual care, the benefits were not uniform across all participants. This underscores the importance of patient screening and tailoring program content to individual needs. Future research should further investigate these considerations to optimize digital health intervention design and implementation.

Introduction

Chronic diseases, including cardiovascular disease, diabetes, dementia, and cancer, are the world’s leading cause of death, accounting for approximately 74% of all-cause deaths []. This alarming situation has led to chronic diseases being identified as a major priority for action by the United Nations and the World Health Organization [].

The etiology of chronic diseases is complex and multifactorial, and the current epidemic is largely attributed to a nutritional transition marked by the adoption of Western lifestyles. This transition, characterized by lower-quality diets, sedentary lifestyles, smoking, and excessive alcohol consumption, is evident in both high-income and low-income countries []. In response to these challenges, public health experts underscore the significance of preventive interventions at individual and population levels to reverse these trends and promote well-being []. However, current preventive and digital health interventions face important limitations. Many digital programs rely on static content delivery, lack personalization, and often fail to sustain user engagement over time. Barriers such as digital literacy, limited interactivity, and reduced adherence further restrict their long-term effectiveness [,]. These gaps create opportunities for innovative approaches. Conversational agents (CAs), through their interactive and adaptive design, can address these limitations by providing real-time feedback, personalized guidance, and continuous engagement, thereby offering a potentially more effective tool for promoting sustainable dietary behavior change [-].

In the digital age, between 20% and 80% of people use the internet to monitor their health and access a variety of resources []. Therefore, digital interventions offer promising ways to change lifestyle-related behaviors, particularly dietary ones. These interventions are defined as products or services that use computer technology to promote behavior change. They can be accessed through various media, including handheld devices, digital platforms, and smartphone apps [].

Previous research [-] has shown that in addition to their possible effect on dietary behaviors, these interventions can lead to significant changes in several areas of health, including weight management [], smoking cessation [,], increased physical activity (PA) [], reduced alcohol consumption [], or self-management of chronic diseases []. Their effectiveness often depends on engagement and ongoing interaction with the target population [].

Several types of digital interventions have already been used to improve dietary behavior. mHealth apps provide tools for self-monitoring and goal setting, while web-based platforms deliver structured educational programs and personalized feedback [,]. Wearable devices have also been used to track dietary intake and PA, thereby supporting behavior changes []. While these tools can deliver short-term improvements, they are often limited by a lack of personalization and interactivity, resulting in declining user engagement over time []. CAs, in contrast, offer a dynamic, interactive interface that simulates human dialogue, providing real-time, context-sensitive guidance and maintaining ongoing engagement. These gaps create opportunities for an innovative approach [-,,].

Based on artificial intelligence, CAs simulate human conversations and provide personalized support to users. Their ability to provide real-time information, adapt to individual needs, and maintain ongoing engagement makes them particularly interesting for health interventions []. In the field of nutrition, chatbots could play a crucial role in providing personalized nutritional advice, helping with meal planning, answering nutrition-related questions, and encouraging healthy eating habits. Their 24/7 availability and ability to process large amounts of information make them potentially powerful tools for supporting lasting changes in dietary behaviors []. However, despite their apparent potential, the true effectiveness of CAs in improving dietary behaviors and preventing diet-related chronic diseases remains largely unexplored. Systematic evaluation of their impact, features, and user acceptability is crucial to determining their value in public health strategies to promote the adoption of healthy dietary behaviors and ultimately prevent diet-related chronic diseases [].

To our knowledge, no systematic review has specifically examined the effectiveness of chatbots on dietary behaviors in the general population.

The primary objective of this mixed systematic review is to assess the effectiveness of interactive CAs designed to improve dietary behaviors. Secondary objectives are to list their basic features, functions, and conversational capabilities and, where possible, also assess their impact on nutritional knowledge, as well as their usability, acceptability, user experience, and engagement.

Methods

Overview

The review was structured using the Preferred Reporting Items for Systematic Reviews (PRISMA) guidelines () []. To ensure transparency and rigor, a completed flow diagram was included in accordance with the PRISMA recommendations.

The systematic review was conducted in several stages: a comprehensive literature search, meticulous study selection, detailed data extraction, rigorous quality assessment, thorough data analysis, and synthesis of findings []. The protocol for this review was registered with the International Prospective Register of Systematic Reviews (PROSPERO) under the identifier CRD42023458561.

Eligibility Criteria

This review adheres to clearly defined eligibility criteria, guided by the Population, Intervention, Comparison, Outcomes, Study Design (PICOS) framework, to ensure a focused and reproducible selection of relevant studies, enhancing the rigor of the systematic review process. The population of interest is the general population, regardless of age, to maximize the scope and generalizability of the findings. The intervention focuses on interactive CAs delivered via any available interactive digital platform, such as smartphones, web applications, or other digital devices, specifically designed to improve dietary behaviors. Both comparator and no comparator studies were considered to ensure inclusivity and to capture a wide range of evidence. The primary outcome of interest is the effect of CAs on dietary behaviors, while secondary outcomes include assessments of nutrition knowledge, usability, acceptability, engagement, and user experience. To leverage the complementary strengths of different research approaches, we included quantitative, qualitative, and mixed methods studies in this review to provide a holistic understanding of the impact of CAs on dietary behavior change [].

We included studies published since 2013 in English, French, or Spanish. This timeframe was chosen based on the findings of Lyzwinski et al [], who reported that 96% of publications on the use of chatbots for lifestyle behavior change were published after 2013. Languages were chosen because they are predominant in scientific literature, and the research team members are fluent in these languages. Dissertations, theses, and relevant websites of recognized nutrition-related organizations were not included in the review due to time limitations and insufficient human resources. We excluded syntheses of knowledge such as systematic reviews, editorials, opinion pieces, conference abstracts, and commentaries.

Search Strategy

The search strategy was developed in collaboration with a librarian (DZM) and is detailed in . This strategy was implemented in 5 electronic bibliographic databases: MEDLINE, CINAHL, Embase, Web of Science, and PsycINFO. The search covered literature published between January 2013 and December 2024. The initial database search was conducted on October 16, 2023, and updated on December 17, 2024. In addition, the reference lists of relevant papers were reviewed to ensure the inclusion of all eligible studies.

Data Collection and Analysis

Study Selection

We conducted the review using the online platform Covidence systematic review software (Veritas Health Innovation Ltd). We imported all references into the tool, and most duplicates were removed automatically. Three reviewers (SA, SMARD, and AB) independently assessed the abstracts and titles of the studies identified by the search strategy after removing duplicates. Relevant studies were selected according to the predefined inclusion criteria. Following this stage, three independent reviewers (SA, SMARD, and AB) screened the full texts to identify eligible studies. A fourth, senior reviewer (SD) was consulted to resolve any disagreements.

Data Extraction and Management

Two reviewers (SA and SMARD) independently extracted data from the included studies. The following information was extracted using an extraction grid: general information (title, authors, country of study, funding, and year of publication); study details (aim, design, inclusion and exclusion criteria, method of randomization, and allocation); study population (age, sex, sample size, and number for analysis); intervention characteristics (type, duration, follow-up points, chatbot name, broadcast platform, language, and interactivity); and outcomes (primary and secondary outcomes, and method of outcome assessment). Disagreements were resolved by consensus or by consultation with a third reviewer. Authors were contacted in case of missing information or ambiguity.

Assessment of Quality in Included Studies

After the data extraction step, 2 independent reviewers assessed the study quality of the included studies using the Mixed Methods Appraisal Tool (MMAT; McGill University) []. In case of disagreement, the reviewers discussed the matter before consulting a third reviewer.

The MMAT assesses the methodological quality of 5 study categories (qualitative, randomized, nonrandomized, quantitative descriptive, and mixed methods). Each study is appraised against 5 criteria, rated as “yes,” “no,” or “can’t tell.” According to the MMAT guidelines, calculating a single overall score is not recommended. However, consistent with common practice in systematic reviews, we summarized the proportion of criteria met and presented the results as stars (from 1 star=20% of criteria met to 5 stars=100%). Higher ratings indicate stronger methodological quality.

Data Synthesis and Analysis

We conducted a descriptive analysis to summarize the extracted data and provide a narrative synthesis of the findings. This included a breakdown of results by subgroups such as age, gender, ethnicity, and geographical location of the population. Through the data synthesis, we described the effectiveness of current CAs; identified key features, functions, and conversational capabilities of successful chatbots; assessed their usability, acceptability, engagement, and user experience; and highlighted limitations and future research directions.

A meta-analysis was not planned because the study designs, populations, interventions, and outcome measures were too different from each other to allow meaningful quantitative pooling of the data.

Deviations From the Registered Protocol

Initially, we decided to exclude conference papers, but during the selection process, we realized that some conference abstracts were showing results. We have included 1 conference abstract []. Despite the initial intention to explore gray literature sources, including dissertations, theses, and relevant organizational websites, this endeavor was not realized due to time constraints and limited human resources.

Results

Study Selection

The first search of the electronic databases was performed in November 2023 (1619 citations) and was followed by an update in December 2024 (581 citations). These 2 searches yielded 2200 citations in total. After 561 duplicates were removed, the titles and abstracts of the remaining 1639 studies were screened, excluding 1519 studies. The remaining 120 studies were read in full and excluded if applicable, leaving a final set of 11 studies. The detailed process of study selection is presented in the PRISMA flow diagram ().

General Characteristics of the Included Studies

The characteristics of the selected studies are summarized in Table S1 in [,-]. The 11 included papers were published between 2013 and 2024 with one of these published in 2024 [], 5 in 2022 [,,,], 1 in 2021 [], 1 in 2020 [,], 2 in 2017 [,], and 1 in 2013 [].

All studies were conducted in high-income countries: the United States (5/11, 46%), the Netherlands (2/11, 18%), Australia (1/11, 9%), the United Kingdom (1/11, 9%), Singapore (1/11, 9%), and the Republic of Korea (1/11, 9%). Most of the studies (6/11, 55%) were randomized trials. The remaining studies incorporated a combination of qualitative (2/11, 18%), mixed methods (1/11, 9%), and pre-post (2/11, 18%) studies. The sample size of the 10 studies that provided this information ranged from 20 to 480 participants. Almost all the studies (10/11, 91%) were performed among adults (mean age ranging from 15.0, SD 0.7 y to 73, SD 5.33 y) while 1 study was conducted among adolescents (mean age of 15.0, SD 0.7 y).

Most of the reviewed interventions (9/11, 82%) included more females than males, and the ethnicity of the included studies that disclosed this characteristic of participants was mainly from the Caucasian (White) or Chinese populations. Four of the studies (4/11, 36%) included participants with obesity and/or overweight as a health condition.

All studies focused on interventions that primarily aimed to change dietary behaviors through CA. The duration of these interventions varied from a single interaction to 12 months. CAs used in interventions are more frequently embodied CAs (7/11, 64%) rather than rule-based CAs (4/11, 36%). Embodied CAs are computer-generated animated humanlike characters that interact with users through verbal and nonverbal behavioral cues []. Rule-based CAs match the user input to a rule pattern and select a predefined answer from a set of responses with the use of pattern-matching algorithms []. The majority of these CAs (8/11, 73%) used avatars, seven of which were female, and one of which allowed users to choose the ethnicity of the avatar.

The interaction patterns of these agents were predominantly text-based (10/11, 91%), with some incorporating voice interactions, figures, images, or video links. A minority (1/11, 9%) used buttons or drop-down lists for predefined responses. The deployment of these agents occurred via digital platforms (7/11, 64%) or social networking applications (3/11, 27%) such as Facebook Messenger or Slack. Several theoretical models and techniques were used in the development of these CAs such as the transtheoretical model (TTM; 7/11, 64%), social cognitive theory (2/11, 18%), self-determination theory (SDT; 1/11, 9%), fuzzy trace theory (1/11, 9%), motivational interviewing (3/11, 27%), shared decision-making (1/11, 9%), persuasion techniques (1/11, 9%), nudging techniques (1/11, 9%), and feedback and rewards (1/11, 9%).

Outcomes of the Included Studies

Dietary Behavior Effects

Overview

This review of studies examining the impact of CAs on dietary behaviors reveals a paucity of consistent results across various outcomes (Table S2 in [,-]).

Consumption or Intake

Studies (4/11, 36%) examining dietary intake have shown varied outcomes across different food groups and intervention approaches. Regarding fruits and vegetables, Bickmore et al [] found that participants in the diet group consumed significantly more daily servings compared to the control group. The combined intervention (PA+diet) also led to improvements, though the diet-only intervention yielded the best results (P=.005). Gardiner et al [] reported a significant increase in fruit consumption among users of CAs compared to the control group participants (μ=2-4 vs μ=2-2; P=.04), although changes in vegetable intake were not statistically significant. In contrast, Kramer et al [] found that CAs were unable to persuade users to modify their fruit and vegetable intake. Maher et al [] demonstrated notable improvements in adherence to a Mediterranean-style diet, including fruits and vegetables, at 6 weeks, with these changes maintained at 12 weeks (mean change from baseline to 12 wk: 5.7; 95% CI 4.2‐7.3; P<.001).

Findings related to protein consumption were less consistent (3/11, 27%). Brust-Renck et al [] observed no significant differences between the intervention group and the control group in the weekly consumption of proteins such as fish or red meat. Similarly, Gardiner et al [] reported no notable changes between groups. While Maher et al [] identified initial improvements in healthy protein intake as part of Mediterranean diet adherence, participants faced difficulties meeting recommended servings by week 6.

For whole grain intake (3/11, 27%), Brust-Renck et al [] and Gardiner et al [] both reported no meaningful differences between the intervention group and the control group. Maher et al [] noted improved adherence to Mediterranean diet guidelines, including whole grains, but participants expressed challenges in maintaining the recommended intake.

Other dietary intakes yielded mixed results (4/11, 36%). Saravanan et al [] observed no significant differences between experimental groups regarding calorie or sugar reduction, but 86% of participants across conditions achieved their goals. Brust-Renck et al [] found no significant changes in sugar and sodium intake as well as other dietary intakes between the intervention and the control groups. Lee et al [] reported a significant 60% decrease in weekly sugar intake from beverages while sodium and caffeine consumption from carbonated and energy drinks did not decrease. Kramer et al [] reported that CA interventions did not influence liquid intake.

Eating Behaviors

Eating behaviors are defined here as a “normal behavior related to eating habits, selecting foods that you eat; culinary preparations and quantities of ingestion” []. In our review, all interventions aiming to improve eating behaviors yielded mixed results, with some studies highlighting limited impact and others demonstrating moderate success (4/11, 36%). Gardiner et al [] found that 69% of participants used CA suggestions to improve eating habits such as consumption of fruits and vegetables, soda, caffeine, snacks, whole grains, red meat, and fish, but this was not significantly different from those relying on patient information sheets (66%). Kramer et al [] identified a significant correlation between competence and eating behaviors (r=–0.38; P=.03), with competence also predicting eating behavioral changes over time (F_1.30=4.30; P=.047; R²=0.13). Pecune et al [] noted that 47% of CA users accepted healthy recipe recommendations, which were healthier than their initial preferences. While participants tended to opt for healthier recipes when the CA included explanations with its recommendations, the results did not reach statistical significance. Finally, Smriti et al [] emphasized the influence of parents on their children’s eating habits, as improved parental feeding behaviors positively shaped those of their children. CAs facilitated reflection and supported families, though challenges related to complex family dynamics were noted. Overall, these findings suggest that while CAs can moderately influence eating behaviors, barriers such as family factors need to be addressed.

Behavioral Intentions

Only 2 (18%) studies examined behavioral intentions. Brust-Renck et al [] reported that behavioral intentions were correlated with self-reported adoption of a healthy diet (weekly consumption of fruits and vegetables, fish, whole grains, sugar, and sodium). Declared intentions regarding nutrition were also linked to participants’ understanding and adherence to fundamental dietary recommendations. In contrast, Dhinagaran et al [] reported that participants expressed no intention of changing their lifestyle, reflecting a lack of readiness to engage with the intervention.

Self-Efficacy

Gardiner et al [] reported no significant differences in self-confidence in eating healthily between participants who used CAs and those relying on patient information sheets.

Adherence to Dietary Recommendations

Studies evaluating adherence to dietary recommendations (3/11, 27%) provide insights into how CAs can influence dietary behaviors. While adherence was mentioned in 3 studies [,,], only Maher et al [] explicitly reported measurable improvements.

Maher et al [] focused on adherence to the Mediterranean diet, demonstrating significant improvements over a 12-week intervention led by a virtual health coach named Paola. Participants increased their Mediterranean diet adherence scores, maintaining these changes throughout the study.

Bickmore et al [] incorporated dietary recommendations from the National Institutes of Health and the National Cancer Institute into their intervention. These guidelines were used to encourage increased fruit and vegetable consumption and to set personalized dietary goals. However, the study did not report specific data measuring participants’ adherence to these recommendations.

Brust-Renck et al [] used American Heart Association (AHA) dietary recommendations to design their intervention, emphasizing gist comprehension to promote heart-healthy behaviors. However, adherence to AHA recommendations was not directly measured, making it difficult to assess the effectiveness of the intervention in changing long-term behaviors.

Main Characteristics of CAs

Across the included studies, 2 main categories of CAs were identified: embodied CAs (ECAs) and rule-based CAs (RBCAs). ECAs (7/11, 64%) were the most common and primarily relied on text-based chat (6/11, 55%), with some incorporating voice-based interactions (4/11, 36%). These were typically deployed through web platforms (6/11, 55%) or dedicated applications (eg, Slack; 1/11, 9%) and frequently featured female avatars (5/11, 46%). In some cases, both male and female avatars (1/11, 9%) were available, or avatars were culturally adapted (eg, African American avatars, 2/11, 18% or avatars representing 3 different ethnicities, 1/11, 9%). In contrast, RBCAs (4/11, 36%) relied on structured interactions through predefined options such as buttons, drop-down menus (1/11, 9%), or scripted responses (1/11, 9%), and were integrated into social media platforms (eg, Facebook Messenger: 2/11, 18% or web-based interfaces: 1/11, 9.1%), often with minimal representations (eg, a robot head: 1/11, 9% or no avatar at all: 1/11, 9%).

From a theoretical perspective, both ECAs and RBCAs commonly drew upon the TTM (7/11, 64%) and social cognitive theory (2/11, 18%), while some incorporated additional frameworks such as SDT (1/11, 9%), fuzzy-trace theory (1/11, 9%), motivational interviewing (3/11, 27%), and shared decision-making (1/11, 9%). Techniques including persuasion (1/11, 9%), nudging (1/11, 9%), feedback (1/11, 9%), and rewards (1/11, 9%) were also reported as strategies to promote engagement and support behavior change. While text-based interactions predominated, some interventions included voice features (4/11, 36%) and visual elements (2/11, 18%) such as figures, images, or videos. However, 2 (18%) studies provided limited or unclear specifications regarding the agents’ modalities (1/11, 9%) [] or theoretical underpinnings (1/11, 9%) [], reflecting heterogeneity in design and implementation.

Secondary Outcomes

Overview

In addition to the dietary behaviors–related primary outcomes discussed earlier, a review of secondary outcomes from included studies revealed a complex yet promising landscape (Table S3 in [,-]).

Nutritional Knowledge

Nutritional knowledge has improved significantly in several studies (5/11, 46%). Brust-Renck et al [] reported enhanced knowledge about energy balance, food labels, fast food, and advertising, with healthier self-reported dietary behaviors associated with greater nutrition knowledge. The study highlighted that users who actively engaged with the tutorial demonstrated a better understanding of the AHA dietary principles. Lee et al [] found that awareness of nutrition labels increased from 64.3% to 92.9%, and nonreaders of nutrition labels decreased from 42.9% to 16.7%. Dhinagaran et al [] highlighted positive feedback on diabetes prevention content, which was considered detailed and informative. However, participants who were already familiar with healthy living found the content less relevant.

Other Lifestyle Behaviors

PA outcomes were reported in 2 (18%) of 11 studies. Bickmore et al [] found that participants in the PA group increased daily walking more rapidly than the control group, although no significant differences in International Physical Activity Questionnaire scores were observed. Brust-Renck et al [] noted improvements in PA-related knowledge and self-reported behaviors. Significant increases in weekly PA were reported by Maher et al [], with gains of 109.8 minutes over 12 weeks.

Stress management was another area of focus. Dhinagaran et al [] highlighted the positive effects of CAs in promoting relaxation techniques, such as mindfulness, while Gardiner et al [] reported a significant reduction in alcohol use for stress relief among the CA intervention participants compared to the control group (P=.03).

Social Support

The impact of CAs on social support outcomes was observed in 1 (9%) study. Kramer et al [] observed no reduction in loneliness among participants, indicating limited influence in this domain.

Motivation

Several studies (3/11, 27%) reported positive effects on motivation to engage in behavior change. Gardiner et al [] found significant progress in advancing stages of change at 6 months, particularly in behaviors related to diet and supplementation, though these effects diminished at 12 months. Saravanan et al [] demonstrated that a memory model, which enables social CAs to recall and reference past interactions to deliver personalized, motivational dialogues based on users’ progress and emotions, significantly boosted motivation, with the greatest increases observed in specific experimental conditions created to evaluate the effects of memory references and run the experiment as a between-subjects design. Smriti et al [] noted that CAs encouraged parents to reflect on their eating habits, motivating them to adopt healthier behaviors that benefit both themselves and their children.

Engagement

User engagement with CAs is less well documented, and this engagement varies between studies (5/11, 46%). Gardiner et al [,] reported a median duration of interactive session with a CA of 13.7 minutes and a median of 6 log-ins over 12 months. Kramer et al [] observed higher engagement levels, with participants logging in to PACO, a web-based eHealth service in which 2 ECAs engage in dialogue with an older adult, an average of 39.97 times and spending most of their time on food diaries (85.45%). In Kramer et al [], engagement tended to decrease over time. In Lee et al [], engagement was classified as active or passive based on whether the data was entered before or after the daily reminder from the chatbot at 8 PM. Only a small percentage (22.5%) of the data was categorized as active engagement, and 71.8% was classified as passive engagement.

User Experience

User experience is well documented (7/11, 63%) and was generally positive, but highlighted areas for improvement. Participants appreciated the likability [,,], ease of use [,,], and tailored recommendations provided by CAs [,]. Explanations accompanying recommendations improved satisfaction and trust []. However, Kramer et al [] reported lower ratings for enjoyment and perceived usefulness, with 94% of participants unwilling to pay for the service. Maher et al [] noted that participants with limited smartphone skills relied on family members for assistance. Other challenges included patronizing tones [,] or unnatural conversation [,], simplistic content [] or need for more information [,], and perceptions of CAs as unrealistic [,]

Feasibility and Usability

Finally, assessments of feasibility and usability (2/11, 18%) were encouraging. Maher et al [] achieved a recruitment target of 30 participants within 6 weeks and a 75% retention rate over 12 weeks, with 70% of participants meeting engagement targets. Kramer et al [] reported usability scores above the midpoint, with esthetics significantly correlating with usability (r=0.44; P=.01). Lee et al [] found a high usability according to the Chatbot Usability Questionnaire. Recruitment was completed in 4 days, and the retention rate at the end of the intervention was 95.2%, with daily participation rates ranging from 83.3% to 100%.

Additional Results

Although not all the studies included reported supplementary results, some notable observations were made and are worth noting (Table S4 in [,-]). Bickmore et al [] found that weight change over 2 months was not significantly different between the intervention and control groups. Kramer et al [] also found no significant changes in quality of life, autonomy, or competence over time, but identified several correlations, including quality of life, autonomy, relatedness, and number of chat messages that were associated with loneliness (although they did not predict loneliness), while esthetics correlated with usability, and enjoyment correlated with perceived usefulness. In addition, perceived usefulness and enjoyment were associated with greater total use time. In contrast, Maher et al [] reported a modest but statistically significant total weight loss of 1.3 kg (95% CI −2.5 to −0.7; P=.01) from baseline to week 12 and a reduction in waist circumference of 2.1 cm (95% CI −3.5 to −0.7; P=.003) over the same period. However, no changes in systolic or diastolic blood pressure were observed. The remaining studies (8/11, 7%) [,,,-,,] did not report additional results beyond the primary and secondary outcomes discussed previously.

Included studies (4/11, 36%) have proposed several recommendations to enhance user experience and engagement with CAs (Table S5 in [,-]). Among those, optimizing message timing to nonworking hours and allowing free-text questions [], along with a broader range of response options [], would enable more personalized interactions [,]. Integrating CAs with popular platforms like Instagram and WhatsApp, alongside a standalone app, would improve accessibility []. Diversifying formats (eg, voice and video) and tailoring content to different age groups and cultural contexts, with translations and adapted to local dietary recommendations, are vital for inclusivity []. Flexible conversation lengths [,], clear recommendations, short-term health goals [], and a more natural voice would make interactions more user-friendly and engaging, empowering individuals to manage their health effectively [].

Critical Appraisal of the Included Studies

The quality of the included studies was assessed using MMAT [].

In our review, the 2 qualitative studies [,] and the quantitative descriptive study were of very good quality (5 stars) [], the majority of the quantitative randomized controlled trials (RCTs; 5/6) [,,,,] were of average quality (3 stars) to very good quality (5 stars) [], although 1 quantitative RCT [] was scored with 1 star because the outcome data were incomplete and the study did not provide sufficient information to determine whether the outcome assessors were blinded to the intervention. In contrast, the quality of mixed method studies varies. One study [] was awarded 2 stars for quality because participants were not representative of the target population, and outcome data were incomplete. While the other study [] has been awarded 4 stars (Table S6 in [,-]).

Discussion

Principal Findings

This mixed systematic review examined the effectiveness of CAs in improving dietary behaviors. Results indicated improvements in fruit and vegetable intake, adherence to the Mediterranean diet, and nutritional knowledge. Some studies also reported benefits related to PA and stress management. However, in the case of stress, effects were observed in multibehavior interventions not specifically focused on nutrition and thus may reflect broader outcomes rather than direct impacts on dietary behavior. Users generally reported positive experiences, particularly regarding goal setting and tailored feedback, which appeared to enhance motivation to adopt healthier eating habits. Nevertheless, challenges such as limited long-term engagement, inconsistent impacts on social support, and heterogeneity in study design highlight the need for further refinement of CA-based interventions.

It is also important to emphasize that while some improvements did not always reach statistical significance, they may still represent clinically meaningful effects, especially in the context of dietary behavior change, where even modest shifts can contribute to long-term health benefits. For example, even minor enhancements, such as slight reductions in weight and waist circumference, have the potential to yield substantial health benefits on a population level.

Comparison With Prior Work

Recent evidence reinforces the value of explicitly integrating behavior change theories into the design of CA interventions. A recent scoping review [] shows that theory-based designs enhance CAs’ effectiveness in promoting healthy behaviors and improve reproducibility, evaluation, and synthesis of findings.

Our review confirms that many CAs were developed using theoretical models, contributing to the observed impacts. Behavior-change frameworks like the TTM, SDT, social cognitive theory, and fuzzy-trace theory provided structure and alignment with key behavioral determinants, including motivation, self-efficacy, and decision-making processes [-]. For instance, interventions grounded in the TTM have consistently shown significant improvements in dietary behaviors, including increased fruit and vegetable intake and reduced fat consumption, across diverse populations [,]. Recent findings confirm the utility of the TTM in supporting transitions between stages of change and preventing chronic diseases by strengthening self-efficacy and tailoring content to users’ readiness to change []. In digital health, the TTM is commonly used to guide behavior change programs by offering stage-matched content, self-assessments, and personalized feedback, with real-time monitoring and adaptive messaging to sustain self-efficacy and support long-term adherence to healthy eating patterns [,]. Similarly, studies leveraging SDT have shown that fostering autonomy, competence, and relatedness significantly enhances intrinsic motivation to adopt and maintain healthy eating habits [,-]. In digital health interventions, SDT is used to design features that support user autonomy, competence, and relatedness []. Recent digital interventions embedding SDT principles report higher user engagement, stronger intentions to change, and greater satisfaction []. Tailoring content to users’ motivational profiles further supports sustained behavior change and improved adherence, particularly in dietary and PA interventions [,].

Interventions based on fuzzy-trace theory improve decision-making and long-term adherence to healthy behaviors by simplifying information into meaningful gist representations [,]. In digital health, this approach aligns content with users’ core values while reducing cognitive load, improving adherence to health recommendations [].

Similarly, in digital contexts, social cognitive theory provides a robust framework emphasizing self-efficacy, observational learning, and reinforcement. Interventions grounded in this theory have improved PA and treatment adherence in primary care and mHealth apps [].

Beyond informing content, these theories strengthen intervention design by providing structured and replicable frameworks for understanding and influencing behavior change. They help identify key behavioral determinants, guide intervention functions, and enable evidence-based techniques tailored to users’ cognitive, emotional, and motivational states [,]. Recent reviews have demonstrated that theory-based digital health interventions outperform those without a theoretical grounding. For CAs, theories also inform interaction strategies, timing, and adaptation mechanisms, which are essential for dynamic, responsive user experiences []. These findings underline the foundational role of theoretical models in designing effective CA interventions that address the complexity of behavior change.

Behavioral and psychological techniques were also innovatively applied in CA designs. Motivational interviewing, used in 2 studies [,], helped CAs elicit “change talk” and address user ambivalence, thereby increasing commitment to dietary goals []. Recent evidence suggests that CAs using motivational interviewing principles can enhance users’ willingness to change, particularly when interactions are designed to promote cooperation and reflective thinking [].

Nudging techniques, as in Pecune et al [], subtly guided users toward healthier food choices by highlighting preferred options without limiting freedom of choice []. In digital health, nudging improves adherence to treatment and dietary choices through adjustments in choice architecture and visual emphasis [,]. Similarly, fuzzy-trace theory supports intuitive, gist-based messaging to improve decision-making and comprehension [].

Persuasion techniques, when applied ethically, strengthen the credibility and emotional resonance of CA interactions and potentially increase their effectiveness [,,,]. Recent studies show that persuasive dialogue, aligned with users’ emotional and cognitive openness, can influence health behaviors []. Theory-based interventions help operationalize such techniques (eg, goal setting, feedback, and self-monitoring) and aligning them with validated mechanisms of action. For instance, social cognitive theory supports reinforcement, modeling, and self-monitoring strategies to improve engagement and outcomes [].

Several studies integrated empathy [,-] and cultural sensitivity [,] to enhance user engagement. Empathy fosters trust, rapport, and promotes positive user experiences, while culturally tailored interventions increase relevance and acceptability among diverse populations [].

Furthermore, the reviewed studies underscored discrepancies between single-component and multicomponent interventions. Single-component interventions, which focus exclusively on dietary behavior, often exhibit limited effects due to their inability to address the multifactorial nature of behavior change.

In contrast, multicomponent interventions, such as those combining dietary guidance with PA promotion or stress management, tended to yield more significant and sustained outcomes. For instance, Bickmore et al [] demonstrated that multifaceted interventions produced substantial improvements in fruit and vegetable intake and PA levels. Similarly, interventions targeting multiple lifestyle domains, such as diet, PA, sleep, and stress management, have shown positive outcomes beyond dietary behaviors alone. For example, studies by Dhinagaran et al [] and Gardiner et al [] reported that CAs promoting stress-reduction strategies, such as deep breathing and mindfulness, were associated with improved stress management and sleep quality, reinforcing the added value of multicomponent approaches in addressing interconnected health behaviors. These findings align with prior reviews suggesting that comprehensive designs better address the complex interplay of factors influencing behavior change []. Such designs are particularly important for addressing co-occurring barriers to change, such as stress and physical inactivity, which often undermine dietary efforts.

A notable strength of CAs is their adaptability to specific populations. For example, Gabby [,], developed for African American women, addressed cultural barriers and systemic inequities, fostering trust and relevance in health care interactions. Similarly, Herman and Ellen [] targeted older adults, providing tailored support for age-related challenges such as accessibility and social isolation. These examples underscore the importance of designing culturally sensitive and demographically tailored CAs. For African American women, cultural relevance, addressing health care distrust, and personalized support for chronic disease prevention are critical. For older adults, interventions must focus on accessibility, cognitive engagement, and reducing isolation. For this population, technology adoption itself can pose a barrier; for example, interventions that are overly complex may discourage participation. Designing intuitive interfaces, including age-appropriate features, and offering support or training could improve adoption and engagement [,]. These findings are consistent with studies emphasizing the pivotal role of trust-building, empathy, and personalized interventions in enhancing digital health outcomes [].

Studies reported improved nutritional knowledge, particularly in understanding food labels and energy balance. For instance, Reyna et al [] highlighted how intuitive, gist-based messages improve user comprehension, while motivational interviewing and nudging techniques enhance user motivation and adherence. However, challenges remain in achieving consistent improvements in self-efficacy and behavioral intentions, emphasizing the need for interventions that comprehensively address psychosocial determinants of behavior. Integrating real-time feedback and goal-setting mechanisms may enhance these dimensions by providing users with actionable insights and reinforcing positive behaviors. In addition, insights from other digital health research suggest that engagement can be reinforced through mechanisms such as reminders, routine follow-ups, occasional face-to-face contact, or even involving family members in the intervention process [,]. These approaches could support long-term adherence.

Future research should prioritize rigorous, long-term studies to evaluate the sustainability and scalability of CA interventions. Expanding inclusion of underrepresented populations, particularly in low-resource settings, is critical for promoting equity and accessibility [,]. Attention should also be paid to the modes of delivery, as interventions were deployed through smartphones, computers, or web-based applications, each with distinct usability and accessibility implications [,]. Hybrid models combining CA-driven support with human coaching hold promise for addressing engagement and trust challenges []. In addition, standardizing outcome measures and leveraging validated theoretical frameworks will be essential for refining CA design and implementation. Ensuring the credibility of content is another critical consideration. CAs should be designed to deliver evidence-based, regularly updated information, ideally developed with input from multidisciplinary teams including clinicians, dietitians, and patients, to guarantee accuracy and personalization []. These advancements are crucial for scaling effective and equitable interventions to support global health initiatives and prevent chronic diseases.

Strengths

This review has several notable strengths. First, to our knowledge, it is the first mixed methods systematic review to specifically examine the effectiveness of CAs in improving dietary behaviors within the general population. Second, by combining quantitative and qualitative evidence, the review provides a comprehensive overview, capturing both the measurable impact of CAs and the user experience. Third, this review systematically documents the theoretical underpinnings, design features, and behavior change techniques embedded in the interventions, offering insights into the mechanisms contributing to their effectiveness. Fourth, rigorous methodological procedures were followed, including duplicate screening, independent quality appraisal, and structured synthesis. These procedures enhance the transparency, reproducibility, and reliability of the findings. Finally, the review identifies knowledge gaps while highlighting promising design strategies, such as theory-based frameworks, personalization, and multicomponent approaches, that could inform the development of more effective, equitable, and scalable CA-based interventions in the future.

Limitations

This review highlights the potential of CAs to promote dietary behavior change while also revealing key challenges that limit the generalizability and impact of the findings. First, a common issue across the included studies was the focus on specific populations, such as African-American women or individuals with high digital literacy, and the fact that most interventions were conducted exclusively in English. This linguistic and cultural narrowness limits the generalizability of findings and may exacerbate existing health disparities by excluding individuals with lower digital literacy or from non-English–speaking backgrounds.

Many studies relied on participants recruited from online platforms or health-focused communities, introducing selection bias, as these individuals were often already motivated to adopt healthier behaviors.

Second, methodological constraints also impacted the robustness of the findings. Self-reported data were commonly used to measure dietary adherence or behavior changes, which, while convenient, are prone to bias and may overestimate the effectiveness of interventions. Small sample sizes further reduced the reliability of results, and the absence of control groups in certain studies made it challenging to isolate the effects of CA from other factors. Moreover, the short duration of many interventions limited insights into the long-term sustainability of observed changes. While some studies demonstrated initial improvements in dietary behaviors or motivation, follow-up data often failed to confirm whether these benefits persisted.

In addition, differences across study designs (eg, RCTs vs pre-post or qualitative studies), target populations (adolescents, adults, or older adults), and types of CAs (ECA vs RBCA) may have contributed to the heterogeneity of findings. Although our review was not structured by these subgroups, recognizing such variability is important for interpreting results and underscores the need for future reviews or meta-analyses to systematically examine subgroup effects.

Third, very few studies thoroughly evaluated engagement with CAs, usability, or feasibility, limiting the ability to assess their true impact in real-world settings. This lack of process-oriented evaluation makes it difficult to understand how users interact with the CA, what features drive sustained use, and whether the interventions are scalable or adaptable to diverse health system contexts.

Fourth, 3 studies incorporated established dietary guidelines into their interventions. However, adherence outcomes were not always quantitatively measured. Only 1 study [] explicitly reported measurable improvements in their Mediterranean diet adherence scores, thereby leaving the impact of these guidelines uncertain.

Finally, an additional limitation of this review is the exclusion of gray literature, including dissertations, theses, and reports from relevant organizations. Due to time constraints and limited human resources, it was not possible to conduct a systematic search of these sources as originally planned. As a result, it is possible that relevant studies were missing. This could limit the comprehensiveness of the review and underrepresent the diversity of CA-based interventions being explored in practice, especially in nonacademic or community settings.

These limitations underscore the need for more comprehensive testing and validation of CA-based interventions. Addressing these limitations through more inclusive participant recruitment, enhanced technological adaptability, rigorous methodologies, and long-term evaluations will be essential for maximizing the potential of CA to promote sustainable dietary behavior change.

Future Directions

To build on the potential of CA while addressing the challenges identified in this review, future research must adopt a multidimensional approach. First, studies should prioritize the design and implementation of larger, longer-term interventions with robust follow-up periods to assess the durability of behavior changes over time. Understanding the characteristics and needs of diverse populations is critical to ensure interventions are culturally sensitive, linguistically accessible, and equitable. Particular attention should be given to underserved and marginalized groups to expand the applicability of findings and maximize public health impact.

Future research should prioritize the development of culturally sensitive and linguistically diverse CAs to ensure greater inclusivity and applicability. Designing interventions in multiple languages and tailoring them to different cultural contexts is essential for reducing inequities and extending the benefits of CA-based dietary interventions to underserved populations in digital health research.

Developing standardized methodologies is another essential step. The adoption of validated outcome measures will enable a more comprehensive assessment of intervention efficacy. These efforts align with the World Health Organization’s Global Strategy for Digital Health 2020‐2025 [], which advocates for rigorous evaluation frameworks to guide digital health innovations.

Advancements in CA design and functionality will also be central to improving their effectiveness. Enhancing natural language processing capabilities, incorporating advanced personalization algorithms, and designing more intuitive user interfaces will make these technologies more engaging and accessible. Future research should explore how to seamlessly integrate CA into existing clinical workflows through pilot programs and cost-effectiveness analyses. By demonstrating real-world applicability, these initiatives can support the scalability and sustainability of CA-based interventions within health systems.

Equitable access must remain a guiding principle in the evolution of CAs. Bridging the digital divide by addressing barriers to technology adoption, such as literacy gaps, affordability, and cultural relevance, will be critical. Leveraging partnerships with community organizations and public health agencies could further ensure that these interventions reach the populations most in need.

By addressing current challenges and embracing opportunities for innovation, future research can enhance the role of CA in promoting dietary behavior change, preventing chronic diseases, and supporting the integration of these tools into mainstream health care. Such advancements will unlock the full potential of CA to drive meaningful and sustainable health outcomes on a global scale.

Conclusions

This review underscores the promising yet mixed potential of CA in promoting dietary behavior change. While notable studies demonstrate improvements in dietary intake, adherence to healthy eating patterns, and nutritional knowledge, others report limited or non-significant differences, highlighting variability in study designs, content delivery, CA types, and individual motivation. Positive user experiences across most interventions suggest the feasibility of using CA in health promotion. However, concerns about user engagement, satisfaction, and perceived usefulness reveal critical areas for refinement.

Future research should address existing limitations, including small sample sizes, short study durations, and methodological inconsistencies, while exploring ways to enhance the relevance and inclusivity of CA for diverse populations. Focus should be placed on addressing the barriers faced by underserved communities, including limited digital literacy, language constraints, and socioeconomic challenges, to ensure equitable access to these interventions. Integrating CA into broader health care systems, improving their design through advanced artificial intelligence–driven personalization, and evaluating their long-term public health impact will be essential for maximizing their effectiveness. By prioritizing these advancements, CA could play a transformative role in nutrition and chronic disease prevention, providing scalable and accessible tools for global health promotion.

Sign up here.

Rosneft said additional pressure on the company’s results came from an increase in ensuring “anti-terror security”.

Rosneft said its revenue dropped by 17.8% in the first nine months of the year to 6.29 trillion roubles.

“The high level of the Bank of Russia’s key interest rate continues to have a significant negative impact on the profit. In addition, non-monetary and one-off factors adversely affected the indicator’s dynamics during the reporting period,” Rosneft said.

Earnings before interest, taxes, depreciation and amortisation (EBITDA) fell for the period by 29.3% to 1.6 trillion roubles.

($1 = 77.4955 roubles)

Our Standards: The Thomson Reuters Trust Principles., opens new tab

“It’s almost impossible for any developer who is now of working age to have grown up without at least some influence from Nintendo,” says Keza MacDonald, author of the forthcoming book Super Nintendo: The Game-Changing Company That Unlocked the Power of Play, based off years of reporting on the company as a games journalist. “It is still, to this day, making games differently from everyone else.”

Indeed Nintendo has largely sidestepped the graphics arms race that has bedeviled both its hardware and software competitors, instead focusing on what Game Boy designer Gunpei Yokoi affectionately termed “withered technology”: Using well-established technology and focusing on making something fun instead. That strategy has also allowed Nintendo to avoid the high costs and constant retraining that are hamstringing its  competitors. 

The Japanese game developer embraced “the principle of finding a playful way to design things that aren’t necessarily at the cutting-edge,” explains MacDonald, who currently writes about gaming for The Guardian. “That’s been a part of Nintendo’s philosophy since before it was even making video games.”

The Japanese company has what MacDonald deems a “slightly conservative” approach, ensuring that it maintains healthy profit margins and builds up large reserves of cash. “Nintendo always operates with an understanding that its next product might not be a hit,” she says. 

Nintendo released the Switch 2, its latest video game console, earlier this year. While a few commentators griped that Nintendo’s latest version was just more powerful (and more expensive) than the last, gamers seem to have flocked to the new device. The company now expects to sell 19 million Switch 2 units by March 2026, the end of its fiscal year. The company reported 1.1 trillion Japanese yen ($7 billion) in revenue between March and September, more than double what it generated the same period a year ago. It also earned 199 billion yen ($1.3 billion) in profit, an 83% jump. Shares are up 46% for 2025 so far.

Nintendo was founded in 1889 as a company making playing cards and eventually moved to making toys in the 1960s. It shifted to video games in the 1970s, and had its first hit with Donkey Kong, developed by Shigeru Miyamoto, who eventually designed beloved franchises like Super Mario and The Legend of Zelda. 

The game industry is known for its churn: Studios expand and contract according to changing demand. Around 10% of developers reported being laid off last year, and over 40% said they felt the effects of layoffs, according to a survey from the Game Developers Conference. “What that does is it robs companies of not just the knowledge, but also the security that helps people do their best work,” MacDonald says. 

Nintendo, on the other hand, has sidestepped this boom and bust cycle. The company revealed earlier this year that its Japan-based employees had an average tenure of 15 years.

“The people who first made Nintendo’s hits are still working at the company,” MacDonald says. “For the last 50 years, these people have been passing down knowledge and training up a new generation of Nintendo creatives.” 

She adds that the company also rejects hierarchy when it comes to design. “It’s not like the oldest guy gets to decide what’s a good idea and what isn’t. Everyone puts ideas in.”

Not all of Nintendo’s experiments work. Take the company’s Wii U console, released in 2012. Unlike its predecessor, the wildly successful Wii, the Wii U was a flop, selling barely 14 million units. Yet Nintendo took some of the design lessons from this failure and put them towards the Nintendo Switch—which, at 154 million units sold, is close to being the top-selling console of all time. 

That’s just one of the things that MacDonald thinks that other companies—and not just those in the gaming industry—can learn from Nintendo. 

“A failed idea is often a step towards the next hit you’re going to have.”

RIO DE JANEIRO, Nov 28 (Reuters) – Brazilian state-run oil firm Petrobras could review some of the 15 wells it plans to drill in the so-called Equatorial Margin, as Brent oil prices are expected to remain low in the coming years, its CEO said on Friday.

Sign up here.

“We had a large set of wells for the Equatorial Margin; some were prioritized, others were, let’s say, deprioritized depending on the price of Brent crude oil,” Petrobras Chief Executive Officer Magda Chambriard said in a press conference.

She did not say how many of the 15 wells could be reviewed.

Petrobras’ cuts will also impact extraordinary dividends to shareholders, Chief Financial Officer Fernando Melgarejo told journalists, saying the possibility of doling out extra cash in the coming years is low.

Despite the cuts, Petrobras expects to maintain its oil production at some 2.6 million or 2.7 million barrels per day until 2034 after ramping it up around 2027, Chambriard said.

These oil production levels represent the peak Petrobras expects to reach in the next five years under its new business plan.

Our Standards: The Thomson Reuters Trust Principles., opens new tab

Theo Leggett
Business correspondent

The issue was discovered after a JetBlue aircraft en-route from Mexico to the United States in October experienced a ‘sudden drop in altitude’.

The plane made an emergency landing, with reports at the time suggesting 15 to 20 people suffered minor injuries.

It’s thought the incident was caused by intense solar radiation, which corrupted data in a computer used to help control the aircraft.

Now action is being taken to prevent further problems. About 6,000 aircraft worldwide are thought to be affected, all of them of the A320 family, which also includes the A319 and A321 models.

According to Airbus, the majority can be fixed with a relatively simple software update. However, some 900 older planes will need replacement computers, and will have to be taken out of service until they can be fixed.