The study design and main outcomes of the RCT have been published previously [9,13]. In brief, the trial (German Clinical Trials Register: DRKS00022923) was a randomized, controlled, open-label, parallel-group study with a 2:1 allocation ratio favoring the intervention group. The 3-month follow-up compared the intervention arm—using mySugr for 3 months—with a control group without a digital health intervention. The key inclusion criteria were a diagnosis of type 1 diabetes, type 2 diabetes, or GDM; regular daily self-monitoring of blood glucose; age≥16 years; and a most recent HbA_1c value <12% (107.6 mmol/mol). The key exclusion criterion was the use of a continuous glucose monitoring (CGM) system. The RCT was conducted in Germany across 41 secondary care practices. For this secondary analysis, only data from individuals with type 1 or type 2 diabetes were used. Additionally, the analysis focused exclusively on HbA_1c as the marker of glycemic control. This secondary analysis was not prespecified.

The app was developed to support individuals with diabetes in the daily management of their condition. It integrates a range of diabetes-related data, including glucose levels, estimated HbA_1c (eHbA_1c) values, and insulin intake, as well as entries on diet, weight, blood pressure, activity levels, and stress. These data can be shared with the diabetes care team before appointments. The app provides users with an overview of their diabetes data through easy-to-interpret reports that use traffic light colors to highlight areas for attention and suggest opportunities for improvement. However, it should be noted that the app does not offer recommendations for therapy adjustments. In addition, it incorporates psychological features such as motivational challenges and positive feedback. A more detailed description of the app is available elsewhere [9].

HbA_1c was used as the marker of glycemic control, assessed at baseline and at the 3-month follow-up visit. HbA_1c levels were measured using venous blood samples.

The difference in the proportion of participants achieving an HbA_1c value ≤6.5%, based on the treatment algorithm of the American Association of Clinical Endocrinologists (AACE) [12], at the 3-month follow-up was analyzed using robust logistic regression. The dependent variable was a dichotomized indicator of whether a participant achieved an HbA_1c value ≤6.5%. The independent variable of interest was group allocation, and the analysis was adjusted for baseline HbA_1c. Only participants with type 1 or type 2 diabetes were included in this analysis, as none of the women with GDM had baseline HbA_1c values above 6.5%. The analysis was conducted for both the intention-to-treat population (ie, all randomized participants) and the per-protocol population. The per-protocol population was defined as participants who met all eligibility criteria, completed the follow-up visit within 42-137 days (ie, within 50% of the planned 84-day study period), used the intervention app on at least 10% of study days (intervention group), did not use any digital diabetes diary (control group), and did not use CGM during the study period—as prespecified in the RCT analysis plan [9]. Missing HbA_1c data in the RCT were imputed using multiple imputation (see [9,13]).

Ethical approval was obtained from the State Chamber of Physicians of Baden-Wuerttemberg (approval number F-2020-121) as the primary vote, as well as from the 13 local State Chambers of the participating study sites. All participants provided written informed consent, which also covered the use of their data for both primary and secondary analyses. Therefore, no additional consent for secondary analyses was required. All RCT data were handled using participation codes (pseudonymization). The list linking participant names to these codes was securely stored at the study sites and destroyed after the study concluded. As a result, only anonymous, deidentified study data were used for all analyses. Participants received a compensation of €50 (US $57.78) for participation in the RCT, which was approved and deemed appropriate by the ethics committee.

RWD was drawn from a user database of existing mySugr users. For this analysis, only users with type 1 or type 2 diabetes from Germany were included, as there is no clinical indication to lower HbA_1c in women with GDM. To align with the RCT’s exclusion criteria, only users who were not using a CGM were selected. To enable analysis of glucose trajectories and comparisons of glucose levels before and after app use, users were required to have uploaded glucose data before initiating app use (defined as baseline, or month – 1) and to have logged at least one blood glucose measurement per month. Glucose data were extracted for 2 cohorts: (1) from baseline (month – 1) up to month 3 of app use, and (2) from baseline up to month 12. The end point of interest was eHbA_1c at various time points, used as a marker of glycemic control.

Propensity score matching was used to select users from the existing user database to enhance the comparability of the RWD with the intervention group of the RCT. This approach was intended to mitigate the lack of randomization and minimize confounding [7]. The following variables were used for direct matching to the intervention group: age (in 10-year intervals), diabetes type, and initial eHbA_1c (categorized in 1% steps from 5% to 10%). In addition, propensity scores were calculated within these directly matched groups based on app usage and other demographic variables. The following covariates were included in the propensity score model: years since diagnosis, self-reported gender, self-reported therapy type, metformin use, number of blood glucose logs, number and the average value of carbohydrate logs, total steps logged, number and the average value of BMI logs, number and the average value of blood pressure logs, number and total duration of exercise logs, number and the average value of basal insulin injections, number and the average value of bolus insulin injections, number of challenges completed in the mySugr app, number of tags and notes added to log entries, number of meal images uploaded, and use of the mySugr bolus calculator. All features were extracted from the first 30 days of mySugr use for both the RCT population and the existing user database. Categorical features were one-hot encoded. To address the class imbalance, the existing user database was randomly undersampled, retaining 20% of the original data. Missing values were imputed using a K-nearest neighbors imputer with the 2 closest neighbors. The data were then standardized, and propensity scores were calculated using a logistic regression model, with membership in the RCT population as the treatment indicator. Matching within the directly matched strata was performed by selecting the 50 closest neighbors for each treated individual based on the logit-transformed output of the classifier, allowing replacement and applying a maximum allowable distance of 0.4. Matching success was assessed using Cohen d values for all included variables.

The same app as in the RCT was used in the observational, real-world study.

Glucose data uploaded from blood glucose meters into the mySugr app were used to calculate the mean glucose for each month over a 12-month period. The eHbA_1c before app use was calculated using the formula by Nathan et al [14].

Changes in mean glucose at month 3 relative to the month before app use were analyzed using a paired, 2-sided t test to assess whether the change differed significantly from 0. Additionally, a 1-sample Wilcoxon signed rank test was conducted to confirm that the rejection of the null hypothesis was not driven by violations of the t test assumptions. All P values were adjusted using the Bonferroni correction to account for multiple testing. To analyze long-term changes, the difference in mean glucose from before app use to month 12 was assessed using the same procedure. Missing data were not imputed. In addition, changes in mean glucose were analyzed separately for people with type 1 and type 2 diabetes, stratified by eHbA_1c values before app use (≤7.5%, 7.6%-8.0%, 8.1%-8.5%, and >8.5%).

P values <.05 were considered statistically significant. All statistical analyses were conducted using R version 4.2.3 (R Foundation; package robustbase for robust logistic regression) and Python version 3.11.4 (Python Foundation), with the following packages: statsmodels version 0.14.0 for all statistical tests, pandas version 2.1.1 for data transformations, and scikit-learn version 1.3.0 for propensity scoring and feature preprocessing.

As part of the onboarding process, mySugr app users can consent to their data being used for research purposes. Only data from users who provided such consent were included in the analyses. The mySugr app adheres to rigorous data privacy and protection standards [15], and only anonymous, deidentified data were used for analysis. Users did not receive any compensation. Given these considerations, ethics committee approval for this specific study was waived (WCG institutional review board tracking number 20251379).

For the RCT data, a total of 342 participants with type 1 or type 2 diabetes were included in the intention-to-treat population, with 225 randomized to the intervention group and 117 to the control group. The per-protocol population consisted of 285 participants with type 1 or type 2 diabetes who completed the RCT in accordance with the study protocol. As shown in Table 1, the majority of participants in the RCT had type 2 diabetes and were receiving insulin therapy.

In the intention-to-treat population, 74 out of 225 (32.9%) participants in the intervention group achieved an HbA_1c value ≤6.5% 3 months after baseline, compared with 31 out of 117 (26.5%) participants in the control group (Multimedia Appendix 2). When controlling for baseline HbA_1c and accounting for the nonnormal distribution of HbA_1c values, the likelihood of achieving an HbA_1c value ≤6.5% at 3 months was nearly twice as high in the intervention group compared with the control group (odds ratio 2.24, 95% CI 1.12-4.47; P=.02). In the per-protocol population, 59 out of 183 (32.2%) participants in the intervention group achieved an HbA_1c value ≤6.5% at 3 months, compared with 26 out of 102 (25.5%) participants in the control group. Similarly, the likelihood of achieving an HbA_1c value ≤6.5% was 2.26 times higher in the intervention group than in the control group (odds ratio 2.26, 95% CI 1.09-4.71; P=.03), when controlling for baseline HbA_1c and accounting for the nonnormal distribution.

For the RWD analysis, a total of 2861 users with at least three months of glucose data were selected, including 348 with type 1 diabetes and 2513 with type 2 diabetes. Propensity score matching was successful, as all demographic variables were comparable between the RCT participants and the RWD cohort (Table 1). Figure S1 in Multimedia Appendix 1 also illustrates the effectiveness of the propensity score matching. At baseline, the majority of individuals with type 1 (223/348, 64.1%) and type 2 diabetes (1645/2513, 65.5%) had an eHbA_1c≤7.5%. Data from 147 users with type 1 diabetes and 1029 users with type 2 diabetes were sufficient for inclusion in the 12-month app use analysis.

Three months after baseline (with month – 1 indicating the month before app use), 1670 out of 2861 (58.37%) users achieved an eHbA_1c value ≤6.5%. Among them, 704 (42.16%) users showed an improvement in eHbA_1c compared with the baseline. By contrast, only 176 out of 2861 (6.15%) users experienced a deterioration, with an eHbA_1c>6.5% at 3 months despite having a baseline eHbA_1c≤6.5%.

Overall, mean glucose levels decreased from 167 mg/dL at baseline (month before app use) to 140.5 mg/dL 3 months after app use began (P<.001; Table 2). This corresponds to a mean reduction of 26.5 mg/dL in the total RWD cohort (Figure 1A, Table 2). Notably, the majority of this reduction occurred during the first month of app use (–22.1 mg/dL) and remained stable thereafter (Figure 1A, Table S1 in Multimedia Appendix 1). Similar glucose trajectories were observed in both people with type 1 (P<.001; Table 2, Figure 1B) and type 2 diabetes (P<.001; Table 2, Figure 1B), with mean reductions of 16.3 mg/dL (95% CI –20.6 to –12.4) and 27.3 mg/dL (95% CI –28.7 to –25.9), respectively, at 3 months postbaseline (Table S1 in Multimedia Appendix 1). Reductions were most pronounced in individuals with a baseline eHbA_1c>8.5%, both in those with type 1 diabetes (–82.2 mg/dL; 95% CI –102.0 to –61.8; Figure 1C, Table S2 in Multimedia Appendix 1) and type 2 diabetes (–104.6 mg/dL; 95% CI –109.1 to –100.3; Figure 1D, Table S3 in Multimedia Appendix 1). Significant reductions in mean glucose over the 3 months following baseline were also observed among users with baseline eHbA_1c>7.5% (type 1: P<.001; type 2: P<.001) and >8.0% (type 1: P<.001; type 2: P<.001). Reductions were also significant (P<.001) among individuals with type 2 diabetes and baseline eHbA_1c≤7.5%. However, no change was observed in individuals with type 1 diabetes and baseline eHbA_1c≤7.5%. Figure S2A-C in Multimedia Appendix 1 illustrates the corresponding mean blood glucose levels before and during the 3 months of app use. Tables S1-S3 in Multimedia Appendix 1 present monthly reductions in mean glucose across the full sample, as well as separately for people with type 1 and type 2 diabetes. Results from the Wilcoxon signed rank test were consistent with those from the paired t test (data not shown).

In the total sample, mean glucose levels significantly decreased after 12 months of app use compared with baseline (P<.001; Figure 2A). In the final month, mean glucose was reduced by 19.8 mg/dL, from 165.9 mg/dL to 146.1 mg/dL (Table 3, Figure 2A). Among individuals with type 1 diabetes, an initial reduction of 14.2 mg/dL (95% CI –19.8 to –9.3 mg/dL) was observed after 1 month of app use. However, this effect diminished over time, with a net reduction of only 4.3 mg/dL after 12 months (Table 3, Figure 2B). By contrast, individuals with type 2 diabetes experienced a sustained and significant reduction in mean glucose of 21.0 mg/dL at 12 months compared with baseline (P<.001; Table 3, Figure 2B). The greatest improvements were observed in users with baseline eHbA_1c>8.5%, with reductions of 73.2 mg/dL in type 1 diabetes (P=.02; Table 3, Figure 2C) and 101.9 mg/dL in type 2 diabetes (P<.001; Table 3, Figure 2D). Figure S3A-C in Multimedia Appendix 1 displays the corresponding mean blood glucose levels before and during the 12 months of app use.

For the cohort with 12 months of data available, monthly reductions in mean blood glucose for the different subgroups are presented in Tables S4-S6 in Multimedia Appendix 1. Among individuals with type 1 diabetes, the greatest numerical reduction in mean glucose compared with baseline was observed at month 1 (Table S4 in Multimedia Appendix 1). For individuals with type 2 diabetes, the largest reduction occurred at month 4 (Table S4 in Multimedia Appendix 1). Similar patterns were confirmed using the Wilcoxon signed rank test (data not shown).