The study is a multi-center, parallel-group, randomized-controlled trial. The research protocol was approved by the Ethics Committee of the Technical University Dresden (MDR ff-EK-421102023_2, Dresden, Germany). The trial was registered at the German Clinical Trials Register (DRKS00025719, DRKS00032395) and conducted in accordance with the declaration of Helsinki (19).

Participants were recruited between February and March 2024 via a study website (https://www.evamebix.de) or directly through the study centers. Interested participants that were recruited online filled in a first screening questionnaire with basic information (age, confirmed T2D diagnosis, availability of a smartphone, contact information) and chose a suitable study center. Afterwards patients were called by trained study personnel for a first screening based on defined inclusion and exclusion criteria (inclusion and exclusion criteria in Table 1) and to schedule a visit at the nearest study center. The target population included patients with suboptimal controlled T2D (HbA1c 7.5%–11%) likely to benefit from lifestyle interventions. Patients with HbA1c >11% or with severe psychological or physical impairments were excluded, because they might require different pharmacological and therapeutical approaches that take priority over a lifestyle-based digital intervention. Inclusion criteria also ensured participants could engage with the digital intervention, as assessed by study personnel during screening. At the baseline visit, eligible patients were informed about the study and gave written informed consent to participate in the study.

Patients were included in the study between February 2 and September 29, 2024. The study was completed on March 4, 2025, with the final 6-month follow-up visit.

Randomization (1:1) of eligible participants into IG or CG was conducted within the electronic case report form (eCRF; secuTrial) using an embedded variance minimization algorithm. The group assignment was triggered upon entry of the participant's baseline HbA1c value by the study personnel. This process allowed concealment of the allocation. The stratification was based on study center and baseline HbA1c level (<8%/≥8%) to ensure balanced distribution across these variables. Patients were randomized 1:1 into either IG or CG.

Although participants were blinded, study staff were not blinded due to their role in assisting with app setup and account activation.

mebix is a CE-certified digital health application for patients with T2D [International Classification of Diseases, Tenth Revision (ICD-10) code E11.-]. The app was preliminarily approved from the BfArM by the time the study was conducted. This approval was based on the results of a separate pilot study that had demonstrated initial evidence of its benefit (20).

mebix aims at helping people with T2D to improve their self-management and lifestyle decisions to improve state of health and quality of life. It is a guideline-based therapy approach that can be used in addition to pharmacological or other therapies initiated by a physician. It is recommended to use mebix at least 6 months and up to 12 months. Within the app, users are guided and reminded through automated messages to set up their goals, track their physical activity and diet as well as physiological parameters. Information on diet and personalized nutrition, physical activity, mental well-being and social aspects of life with diabetes are provided through interactive educational multimedia content. A more detailed overview of the therapeutic approach was prescribed previously (20). For the control group, a placebo version of the app was used that was based on the original app but without the educational content; it only allowed tracking of certain aspects (e.g., nutrition, physical activity).

After the inclusion procedure (screening, randomization) at the assigned study center, the baseline measurements were assessed by the study personnel. Furthermore, patients received a tablet in order to complete patient-reported outcome measures (PROMs). They accessed the online questionnaires through personalized links provided via the secuTrial system.

Afterwards, participants received access either to the mebix app or the placebo app. After study completion, the CG was offered to also use mebix for 6 months. Participants received a compensation of €200-Amazon voucher for their study visits.

After the initial study visit for the baseline assessment, participants visited the study centers again after 3 and 6 months to conduct the follow-up visits. After the 6-months visit, patients were asked to participate in the extension of this trial, that is currently ongoing and aims at assessing the long-term effects of mebix. This extension of the study consists of an additional assessment at the study site 12 months after the original baseline visit.

The primary endpoint of interest is the change in HbA1c levels from baseline (T0) to 6 months. HbA1c levels were measured at the study centers by trained study personnel, following standardized procedures.

Secondary endpoints of interest included the change in body weight (in %) and diabetes-related distress from baseline to 6 months. Body weight was measured during the study visits by study personnel. As a patient-reported outcome, the diabetes-related distress was assessed via online questionnaires using the validated Problem Areas in Diabetes (PAID) scale (21, 22). It consists of 20 questions, each answered on a scale from 0 (not a problem) to 4 (serious problem), indicating the extent to which the items are currently a problem in the patient's life. The final score was calculated by adding up the answers over all questions and then multiplying them by 1.25. This yielded a total score of 100 with results above 40 indicating severe diabetes distress (23).

Further exploratory endpoints included self-efficacy assessed with the Short Scale for Measuring General Self-efficacy Beliefs (German: Allgemeine Selbstwirksamkeits Kurzskala) (24), quality of life assessed with the EQ-5D-5L (25), well-being assessed with the WHO-5 Well-Being Index (26, 27), empowerment, assessed with the Diabetes Empowerment Scale (DES-SF) (28), self-management, assessed with the Summary of Diabetes Self-Care Activities (SDSCA) (29) or depression severity assessed with the Patient Health Questionnaire-9 (PHQ-9) (30, 31).

Statistical analyses were performed using R version 4.3.3 within Rstudio version 2024.02.29. Baseline characteristics for all endpoints are presented using descriptive statistics.

The primary confirmatory analysis was based on the intention-to-treat (ITT) approach, including all randomized participants regardless of protocol compliance. To address confounding from major changes in antidiabetic medication, a modified ITT approach was applied for the primary endpoint whereby outcome data were censored from the time of medication adjustment onward. This allowed for full inclusion of participants while limiting bias introduced by pharmacological effects unrelated to the intervention.

Missing data were handled with multiple imputations (n = 100) using the R function rbmi version 1.4.1 and the implemented “copy increments to reference” (CIR) method, with the control group defined as the reference group (34–36). This method assumes that participants in the IG benefited from the treatment up to the point of dropout. After dropping out, their outcomes are assumed to follow the same trajectory as those in the CG (37). As a result, monotone missing values in the IG were imputed accordingly, while interim missing data were estimated based on the IG's trajectory. For the CG, monotone and interim missing data were imputed based on parameters derived from the CG itself. All of the covariates that were part of the regression model were also used as covariates for imputation.

Differences in endpoints at 3 and 6 months were calculated after imputing the raw values at each time point. Linear mixed models were fitted to each of the resulting imputed data sets and pooled using a multivariate Wald test using the R package mitml version 0.4–5 (38, 39). The independent variables of these models were the within-group factor time (3 and 6 months) and the between-group factor group (IG and CG); included in the model as interacting factors.

To control for potential confounding variables, the baseline HbA1c (included as an interacting factor with time in the model for the primary endpoint), the study center and gender were included. Moreover, a random intercept for each participant was included. Regarding secondary and exploratory endpoints, the model further comprised the baseline value of the respective endpoint as an additional confounding variable.

The results, i.e., mean (M), standard deviation (SD) and 95% confidence intervals (CI) are thus adjusted for all included covariates.

Finally, hypotheses were tested using a priori defined contrasts or one-sided one-sample t-tests, using the R package emmeans version 1.10.1. Due to multiple testing, adjustments for multiplicity were performed using the Benjamini-Hochberg false discovery rate (FDR) correction, applied separately to hypotheses regarding the primary and secondary endpoints. As an additional effect size measure, Cohens' d and the corresponding 95% CI were reported based on the t values from contrasts and one-sample t tests.

A per-protocol (PP) approach was used as a second analysis. Participants who did not comply with the study protocol were excluded, i.e., participants who dropped out or no longer fulfilled inclusion criteria as well as participants in the IG who did not use mebix at all.

Between the 2nd February 2024 and the 29th September 2024, 153 patients were included in the study and randomized either into the IG or the CG.

Overall, 60.1% of participants were male. At baseline the average age was 59.1 years ±9.9 (M ± SD) and the average baseline HbA1c value 8.56 percentage points ±0.96 (M ± SD). Tables 2, 3 give an overview of the baseline demographics and the endpoint characteristics, showing that the groups were comparable. For the secondary and exploratory endpoints, baseline data is not available for the whole randomized sample because one patient of the CG did not fill in any baseline questionnaire. Additional missing values are attributable to patients either omitting responses to individual questions or failing to complete entire questionnaires.

Figure 1 shows the participant flow, including reasons for exclusion from the PP analysis (some participants fell into more than one category for exclusion, see Supplementary Information S1). The most common reason to exclude patients was dropout/loss to follow-up, mostly occurring in the first three months of the study. When comparing the ITT and the PP sample, patients in the PP are more often female and tend to have a lower baseline weight (and BMI) (see Supplementary Information S2).

Based on the ITT analysis, participants in the IG reduced their HbA1c levels an average of −0.80 percentage points (95% CI: [−1.14, −0.46]) after 6 months. This confirmed the hypothesis that mebix leads to a relevant reduction in HbA1c levels of at least 0.3%, with a moderate to large effect size [t(136) = −4.632, p_FDR = 0.004, d = −0.79 [−1.14, −0.44]]. Comparable results were observed in the PP analysis, supporting the robustness of the findings [mean HbA1c reduction: −0.89 percentage points, 95% CI: [−1.17, −0.61], t(93) = −6.327, p_FDR < 0.001, d = −1.31 [−1.76, −0.86]].

While the ANCOVA in the ITT sample did not yield significant main or interaction effects for time and group, the factor group was significant in the PP analysis (Supplementary Information S4). Across analyses, baseline HbA1c was a significant covariate, indicating greater reductions among participants with higher baseline HbA1c levels. Accordingly, the ITT analysis showed that HbA1c reductions were on average 0.38 percentage points greater in the IG than the CG [95% CI: [−0.75, −0.01]; t(128) = −2.010, p_FDR = 0.047, d = −0.35 [−0.70, 0.00]]. The PP analysis confirmed this difference [mean difference: −0.62 [−1.10, −0,15], t(92) = −2.611, p_FDR = 0.011, d = −0.54 [−0.96, −0.13]]. Reductions continued from 3 to 6 months in the CIG, while the CG stagnated or even slightly increased (Figure 2). Overall, significantly more patients in the IG achieved treatment goals, including a relevant HbA1c reduction of at least 0.3% or levels below 7% (Supplementary Information S3).

Overall, both the ITT and the PP analyses support the effectiveness of mebix regarding the glycemic control of people with T2D.

In the ITT-based ANCOVA for weight change, the main factor time was significant, indicating that weight changed over the study period (Supplementary Information S4). For diabetes-related distress, only the baseline level of diabetes-related distress was a significant covariate, suggesting that observed changes were largely dependent on participants' initial distress level (Supplementary Information S4). Comparable results were obtained in the PP analyses, with the additional finding that the main factor group reached significance for diabetes-related distress, indicating favorable results in the IG compared to the CG (Supplementary Information S4).

The one-sample t-tests defined a priori confirmed that participants using mebix showed significant improvements in both secondary endpoints after 6 months, regardless of analysis population. While these improvements were greater in the IG than the CG, they were not significant (see Table 4). Taken together, the consistency across ITT and PP analyses supports the robustness of these findings.

Further exploratory endpoints showed positive developments in the IG, with greater improvements than the CG for most outcomes. Detailed results are summarized in the Supplementary Information S5,S6. Apart from predefined exploratory endpoints, the majority of IG participants used mebix throughout the study period, indicating good adherence and acceptance of the intervention (see Figure 1).

A key strength of the present study is the high adherence observed in the IG, also in comparison to other DiGA trials (41–43). Most participants (75%) used the mebix app during the whole study period. This high level of adherence is also reflected in the user experience ratings, particularly for perspicuity, attractiveness and dependability (Supplementary Information S6). In contrast, dropout in the control group (CG) was three times higher (n = 21; 27.3%), even without accounting for participants who never registered for the placebo app. This suggests lower acceptance and satisfaction with the CG intervention, leading to higher discontinuation.

This study further addresses a common criticism in the body of evidence of DiGA i.e., is the lack of blinding and an active control group. It is the first diabetes DiGA trial to implement a placebo app (44, 45). Unfortunately, the implementation of such an app has not proven to be straightforward. Several participants in the CG were aware of their treatment allocation and never used the app. It is likely that participants informed themselves about mebix and recognized that the placebo app lacked relevant components. This highlights the difficulty of maintaining blinding in digital intervention studies.

A potential limitation is the selection bias inherent in the study population. Participants were likely already interested in mebix and digital interventions for diabetes management which may limit generalizability to the broader population of patients with T2D. Nonetheless, the baseline characteristics of the study sample align with the average mebix user (46), suggesting that the findings are applicable to real-world users who are motivated to engage with digital interventions.

Another challenge in evaluating digital intervention for diabetes is the influence of medication. A continuous adaption of medication is in line with the development of the disease of T2D patients and reflects the standard of care (9). Future studies and analysis should consider methods to account for the complex interplay of diabetes medication and the studied intervention, for example adjusting outcomes with medical effect scores to more accurately isolate the intervention's effects (47).

Lastly, exploratory outcomes may overlap conceptually (e.g., self-efficacy and empowerment, self-management and well-being). However, since these endpoints were considered exploratory, no formal multicollinearity testing was performed, and results should be interpreted accordingly.