We have analyzed data from the HAPPY (Healthy eating using APP technologY) trial, previously described in detail [15]. The main results from the trial have been published [14]. In brief, this was a 2-armed randomized controlled trial comprising women and men with type 2 diabetes that were continuously recruited within 5 primary health care centrs in Stockholm, Sweden. Initial information about the study was provided by health care personnel at their routine health care visit. Patients interested in participating were contacted by study personnel to receive more detailed information. Those who agreed to participate were scheduled for a physical baseline meeting with study personnel. Participants were randomized 1:1 to a 12-week app-based dietary education intervention in addition to regular care, or a control group receiving regular care only. The primary aim of the trial was to examine the effectiveness of the intervention on dietary intake and cardiometabolic risk markers. Study inclusion criteria were: having type 2 diabetes, ≥18 years of age, being able to read and understand Swedish, and having access to and being able to use a smartphone with a personal e-identification. No exclusion criteria were applied. The data collection began in January 2019 and was finalized in August 2023. It was temporary paused during 2020 due to the COVID-19 pandemic. This study is reported according to the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile HEalth Applications and onLine TeleHealth) checklist, which is developed for eHealth or mHealth interventions (Checklist 1).

The trial was conducted in accordance with the Declaration of Helsinki, and approved by the Ethics Committee of the Regional Ethical Review Board, Stockholm, Sweden (2018/652-31; 2018/1094‐32; 2018/2393‐32; 2020‐00591; 2020‐07005; and 2022-02557-02). It was also registered at ClinicalTrials.gov (NCT03784612). All study participants received oral and written information about the study before giving their written consent before the study started. The original informed consent allows secondary analysis of data without additional consent. Participants received no compensation for participation in the study and all data were anonymized after data collection.

The intervention group received the app at study start and the control group received the app at the 3-month follow-up. All participants responded to a web-based lifestyle questionnaire at baseline and again after 3 months (directly after completing the dietary education in the intervention group), 6 months (directly after completing the dietary education in the control group), and after 12-months of follow-up. The baseline questionnaire was filled out before participants were informed about their group allocation. We also measured weight (kg), height (cm), waist circumference (cm), body composition, blood pressure, serum levels of glycated haemoglobin (HbA_1c), and lipids at baseline, and after 3- and 6 months of follow-up.

In total, 133 individuals agreed to participate in the trial. Of these, 68 were randomized to the intervention and 65 to the control group. One participant randomized to the intervention group and 3 participants randomized to the control group dropped out before study start, ie, before knowing their group allocation and before providing any baseline data. In total 119 participants, 65 in the intervention group and 54 in the control group, downloaded the app.

In this study, we addressed change in diet quality, dietary intake, and clinical outcomes from baseline to directly after having completed the dietary education, hereafter written as “short-term” follow-up. Furthermore, we investigated the association on diet quality and dietary intake from baseline to 9 months after completing the dietary education for the intervention group and 6 months after completion in the control group, from here on written as “long-term” follow-up.

The smartphone delivered dietary education focused on the overall diet in accordance with current evidence-based guidelines and the Swedish national dietary recommendations [16]. The app design included the health belief model [17], stages of change model [18], and social cognitive theory [19]. We further included behavior changes techniques such as educational information, goal setting, self-monitoring, feedback, and performance [20]. Participants were encouraged to integrate with the app daily, although it was not a requirement to do so.

The app has six different features, which included (1) educational information, (2) task introduction including a self-set goal of the week, (3) healthy recipes, (4) short fun facts or practical advice, (5) task reminder as a push notification, and (6) task evaluation “how did it go?” with the weekly task. A new topic was introduced each week, for example, “Healthy food patterns” or “Vegetable intake.” Figure 1 shows the 12 topics and gives examples of screens (here translated from Swedish to English) from the app. Each week followed the same activity schedule and included either 11 (weeks 3‐7 and weeks 10‐12) or 12 (weeks 1, 2, 8, and 9) activities depending on whether there were 4 or 5 recipes. The full 12-week course comprised 136 activities.

Users were able to follow their progress and track their percentage of completed activities in the app (see Figure 1). Before an activity is completed, 0% is visualized in the green circle, when all activities are completed for the day, the circle fills up and shows 100% completed activities. Uncompleted tasks were moved to the next day until marked as completed by the user. Information texts, advice, recipes etc, were saved in the app and could be accessed at any time during the study.

For the purposes of this study, data on completed activities was extracted from the app. We calculated an overall user engagement score in percent by summarizing all completed activities by a participant divided by the total number of activities. We also calculated the percentage of completed activities within each activity type. Based on the distribution of the overall engagement score, participants were categorized into groups of low (<50%), moderate (50‐99.9%), and high (100%) user engagement.

Participants dietary intake were assessed using a validated 95-items semiquantitative food frequency questionnaire (FFQ) [21,22] at baseline and after 3, 6 and 12 months of follow-up. Information on how often each item was consumed (eg once per day, week, or month) was reported by the participants. We obtained standard portions sized from the Swedish Food Composition Database from the National Food Agency [23] to calculate average daily intakes in grams of foods and beverages or grams or micrograms for nutrients.

We assessed overall dietary quality using a Nordic Nutrition Recommendations (NNR)-score based on the food-based dietary guidelines from the latest edition of the NNR 2023 [24]. We have described the NNR-score in detail previously [14]. The NNR-score components and the recommended intakes are presented in Multimedia Appendix 1 (see Table S1). In brief, our NNR-score was calculated based on 10 food based dietary guidelines presented in the NNR including vegetables, fruit and berries, whole grains (cereals), legumes, nuts and seeds, fish and seafood, red meat, vegetable oils, sweets including sugar-sweetened beverages (excluding salty snacks), and alcohol. Proportion scores ranged from 0‐3 points for each dietary guideline, with adherence complying to the recommendation yielding 3 points. This resulted in a total NNR score from 0 to 30 points. Furthermore, we examined dietary variables corresponding to the weekly topics of the dietary course (Figure 1).

Background characteristics included age, sex (female and male), education level (≤12 years and >12 years), diabetes duration (<1 year, 1‐5 y, and >5 years), smoking status (never or former, and current), self-reported medication for hypertension, hyperlipidaemia, diabetes, and physical activity. Physical activity was assessed by 2 questions asking about time spent exercising and time spent performing daily activities, for example walking, cycling, or gardening, during a typical week [25]. A dichotomous variable (<150 min/week and ≥150 min/week) was created according to current physical activity recommendations [26]. We measured weight (kg), height (cm), and waist circumference (cm) and calculated BMI (kg/m²). BMI was further categorized as normal weight (<25 kg/m²), overweight (25.0‐29.9 kg/m²), and obesity (≥30.0 kg/m²) [27]. A dichotomous variable for waist circumference of low risk (<88 cm for women and<102 cm for men) and high risk (≥88 cm for women and ≥102 cm for men) for disease was also created [27]. We included baseline fasting HbA_1c (mmol/mol) as well as a dichotomous variable of HbA_1c (<52 mmol/mol and ≥52 mmol/mol), which is a general target value in the treatment of type 2 diabetes in Sweden [28].

Descriptive baseline characteristics were presented as mean (SD) continuous variables, as numbers (n), and percentages (%) for categorical variables. User engagement with the smartphone app was assessed by the proportion (%) of activities completed during the intervention period. We conducted analysis of variance to compare differences in continuous variables between user engagement groups, and χ² tests for categorical data. Aspects of user engagement were also explored.

Paired t tests was used to compare mean changes in diet quality, dietary intake, and clinical markers within user engagement groups. We then fitted linear regression models to analyze differences in changes from baseline to the short-term and from baseline to the long-term follow-up between user engagement groups, with the low user engagement group used as the reference category. Long-term follow-up was not addressed for clinical markers, as these were not measured at the 12-month follow-up. Models were adjusted for baseline values of each dietary variable to account for differences in baseline intakes [29]. Potential confounders were identified using a Directed Acyclic Graph (DAG; see Figure S1 in Multimedia Appendix 1), and we also controlled for age, sex, intervention allocation, education level, BMI, diabetes duration, and HbA_1c level. Mean differences in changes for short-term and long-term intakes between groups, represented by β-coefficients with 95% CI, were calculated. Normality of the resulting residuals of each model was visually assessed using histograms with normal curve overlay and Q-Q (quantile-quantile) plots. All were found approximately normally distributed except for the model of sugar-sweetened beverages. This was because almost no one reported intake of sugar-sweetened beverages, which created no variation and thus a violated model assumption. Therefore, we did not run statistical models to analyse difference in change for this variable. Statistical analyses were performed using Stata (version 17.0; Stata Corporation), with the statistical significance level set at P<.05.

Our results showed that user engagement of a 12-week app-based dietary education targeting people with type 2 diabetes was high. Those with the highest user engagement followed the NNR guidelines to a greater extent and had higher intake of fruit and vegetables already at baseline. High user engagement was associated with beneficial changes in whole grain intake in the short term, that is, directly following the intervention. We found no associations between user engagement and cardiometabolic risk markers. Long-term participants with moderate user engagement continued to report an increase in whole grain intake, along with additional beneficial changes in fiber intake and an increased intake of carbohydrates 6 to 9 months postintervention in comparison to participants with low user engagement. Interestingly, the moderate group had a higher intake of total energy, saturated fat, and sodium at the long-term follow-up, resulting in a significant difference in change compared with the low user engagement group. Thus, the low user engagement group had a comparatively better diet in terms of saturated fat and sodium intake.

Despite the importance of addressing user engagement, studies commonly do not mention this when presenting their results [9,30-32]. The comparison of studies is further complicated by the lack of a standardized definition. Sittig et al [33] reported high user engagement in a 9-week app-based lifestyle intervention, with participants completing, 75.1% of 177 activities, similar to our 77.1%. Their app, like ours, included written educational content and recipes, but also short educational videos. The length of their intervention was also relatively similar, 9 vs 12 weeks, respectively. In another study by Mummah et al [34], 75% (51/68) of participants engaged at least once with their app targeting vegetable consumption during 8 weeks. Low user engagement was reported by Alonso-Domínguez et al [35], with average app use being only 35 days (corresponding to 38.9% of 90 days) during a 3-month multicomponent diet intervention. Data on user-engagement for longer durations are limited, but Lim et al [36] showed that 62% (61/99) of participants used at least one app feature on at least 75% of days during 6 months. Among motivated participants, maintaining high user engagement for longer time periods may be feasible.

Lavikainen et al [12] identified 4 user engagement patterns during a 12-month app-based lifestyle intervention. They found that 46.9% (904/1926) discontinued use, 38% (731/1926) used the app weekly, 10.8% (208/1926) twice a week, and 4.3% (83/1926) daily. Similar to our findings, high user engagement was associated with improved diet quality, and scoring high on a Healthy Diet Index at baseline increased the odds of belonging to a high user engagement group. This suggests that those motivated to use the app already have a healthier diet. Furthermore, Mummah et al [34] showed that baseline vegetable consumption was a moderator of effect and the effect of their intervention increased with baseline consumption. Hendrie et al [13] showed that baseline vegetable consumption and app use were the best predictors of improved variety and intake of vegetables. Interestingly in this case, those with lower baseline intakes increased their variety most and participants with the highest app use had almost doubled their intake compared to participants with the lowest use. Hence, app effectiveness may relate to both app-use and participant characteristics.

We observed a decrease in completed activities over time for the moderate and low user engagement groups, similar to previous mHealth interventions [37-39]. Nevertheless, the majority, over 70%, of our participants completed at least one activity every week. Somewhat fewer participants engaged with the app during weeks targeting sugar intake and slow and fast carbohydrates. One of the most common topics in diabetes management is carbohydrate calculation [40] and potentially our participants felt confident in this, resulting in less engagement. This is in contrast to previous research showing that user engagement can be improved by focusing on health problems or specific needs [10].

Future studies need to investigate reasons for why people stop engaging with apps. Declining app use has been identified as an early indicator of disengagement in a digital lifestyle intervention, highlighting the importance of closely monitoring and support users during the initial phase [12]. Components to increase user engagement in apps, including commercial apps, include gamification (eg, goal setting and quizzes), a more personal approach to app content, such as individualized and tailored reminders, automated data collection, and human-like features [4,41]. A user-friendly design, culturally and personally tailored content, and caregiver support have also been found to play a role [42]. In our app-design, we used gamification elements, such as allowing the user to set a personal goal for the week and progress-track the percentage of activities that had been completed each day, which has been shown to serve positive reinforcement for behavior changes [42]. This, hopefully, encouraged users to be more active in their own progress. However, our app did not include personalized reminders, for example, through timing where the user can set customized reminders based on their needs, which might have increased engagement.

Previous app experiences and social support within apps, such as networking and experience sharing, have been shown to strengthen user engagement, but while social support may influence behavior change for some, it can also introduce negative social competition [42]. We did not include any social elements in our app. In a qualitative study, people with type 2 diabetes preferred app features that supported them to make informed and independent decisions about their own care, rather than following explicit directives [41]. Factors that could make our app more personal and, thus, more engaging could include allowing participants to choose specific activities in the app, such as only receiving educational information, or receiving recipes suitable for their preferences, such as a vegetarian diet.

A strength of this study is that we automatically collected detailed data on app usage. However, we do not know how much time users spent with the app. It is a limitation that we did not explore why user engagement dropped over time. The inclusion of both women and men is another strength, especially since men are usually more challenging to recruit in interventions [43]. The sex distribution and other characteristics, such as age, BMI, and HbA_1c levels, in our study are similar to the average adult population with type 2 diabetes in Swedish primary care [44]. Among our participants, 84.2% were overweight or obese, compared to 82.7% in primary care. However, our participants were slightly younger (63.2 years vs 68.7 years) and had a lower HbA_1c (47.7 mmol/mol vs 52.0 mmol/mol). The fact that participants were recruited in primary care and from areas with relatively high socioeconomic status, and that the app was only available in Swedish, limits the generalizability of our results to other socioeconomic groups.

Additional strengths include use of a validated FFQ [21,22] and addressing both short-term and long-term dietary intake. This allowed us to examine if changes were maintained over time. Furthermore, anthropometrics and cardiometabolic risk markers were objectively measured, but the latter only at short-term follow-up. Different long-term follow-up intervals between participants in the two groups is also a limitation. In addition, power was calculated based on the primary outcome of the trial [15]. The small sample size in this study may have been underpowered to detect statistical significance in secondary analyses. Finally, social desirability bias and recall bias may have influenced reporting in the FFQ. For example, foods considered healthier may be over-reported and unhealthy foods more under-reported [45]. However, we believe this is likely to be random and not related to a specific user engagement group.

In our study, engagement with a smartphone-based dietary education was high, suggesting that an app targeting dietary habits is feasible to use. Those with high user engagement maintained healthier dietary behaviors over time, compared to those with low user engagement. Future studies should address the challenge of designing targeted mHealth solutions to engage individuals in maintaining app-usage during the full extent of interventions.