Identifying optimal diets for individuals with type 2 diabetes (T2D) and prediabetes remains challenging [1]. The World Health Organization defines mobile health (mHealth) as medical and public health practices supported by mobile devices, which have become increasingly prevalent [2]. Together, these developments highlight the growing potential of digital tools to support personalized diabetes management.

Advances in artificial intelligence (AI) now enable the processing of large datasets to deliver personalized nutrition advice that aligns with individual values and preferences, potentially supporting sustained patient engagement in healthy dietary choices [3,4]. However, despite the constant and rapid evolution of diabetes-related mobile apps, few offer adaptive, evidence-based nutrition guidance for patients with T2D in Germany [5-7].

Given the significant interindividual variability in postprandial glucose responses, generic recommendations are insufficient. Integrating data from continuous glucose monitoring (CGM) and user preferences allows for more tailored interventions, using algorithms powered by machine learning and AI to enhance diabetes self-management [6,8-11].

Within the European H2020 “PROTEIN: PeRsOnalized nutriTion for hEalthy living” project (2019-2022), we aimed to develop and evaluate a personalized nutrition and exercise mobile app to support daily T2D management [12]. The PROTEIN project involved 20 partners from 11 European countries, focusing on health promotion through multidisciplinary collaboration. Our exploratory pilot study specifically assessed whether the PROTEIN app could improve dietary compliance and glucose management in T2D, using time in range (TIR) as the primary outcome - a metric not commonly used as a primary endpoint in similar randomized controlled trials (RCTs) [7]. TIR is the proportion of time that a person spends within a target range of glucose levels, usually between 70 and 180 mg/dL, as predetermined by a health practitioner or the person with diabetes [13].

Following Germany’s 2019 digital health reform, only approved apps can be prescribed for diabetes management, ensuring patient safety and reliability [14]. Only 6 of 59 mHealth apps have been approved for people with diabetes [15]. The PROTEIN trial provides prospective, controlled, long-term evidence on AI-driven mobile apps that deliver personalized recommendations using CGM data, user interaction, and preferences. This adaptive AI approach addressing engagement and patient engagement challenges is novel in digital therapeutics.

To our knowledge, this adaptive approach has not been reported in previous studies [16-19]. Therefore, our exploratory pilot study aimed to automatically customize meal plans and physical activity (PA) according to user phenotype and food preferences by integrating CGM data of patients with T2D. The primary objective was to increase the TIR of the study participants by 5% while improving eating behaviors.

Expanding upon existing research findings, we aimed to address the following questions:

Our study addressed two research questions: whether the PROTEIN app increases TIR and whether it affects body weight. The results indicate a modest improvement in TIR (4.2%) during app use, with no change in weight. In summary, our data indicated that the TIR improved with the use of the PROTEIN app for only an average of 12 days, but this fell short of the expected 5% improvement and of the expected 12-week usage period. Unexpectedly, we did not find a significant correlation between the duration of app use and the improvement of the TIR, which could be due to the small sample size and narrow range of usage durations. Despite the high rates of noncompliance in the intervention phase, the TIR improved in parallel with the use of the mobile nutrition app, suggesting a potential signal that a tailored recommendation approach delivered with wearables through a mobile app may be associated with small increases in TIR and may encourage greater concordance with dietary recommendations. However, this observation is exploratory and cannot be distinguished from selection bias or confounding, and any effect on dietary concordance remains uncertain. Our study, therefore, also illustrates the need to balance technological capabilities with user experience since using an app with many features and wearables is time-consuming, contributing to participant fatigue and dropout.

Overall, participants who took part in the intervention with valid CGM data (n=21) had an improvement of 4.2% in their TIR while they were using the PROTEIN app, suggesting that even though they did not follow the recommendations every day, there seems to be a degree of agreement and engagement while using the PROTEIN system [35,37]. It is noteworthy that the outcome relies on the adequate resources of the participants and may be influenced by different factors, including family network, profession, and socio-economic status.

Although the 4.2% increase in TIR reached statistical significance, its clinical relevance remains uncertain, because of the high baseline TIR (mean >70%) and the lack of change in HbA_1c, fasting glucose, and weight. International consensus recommends achieving and maintaining TIR ≥70% and states that a 5% change in TIR is strongly associated with a 0.5% decrease in HbA_1c, which is a clinically significant reduction. We therefore interpret the observed TIR change as a modest signal of potential benefit that requires confirmation in larger, adequately powered studies.

We anticipated that participants would use the PROTEIN app for a full 12 weeks; however, this was not achieved by the majority. This was largely due to the app being perceived as not user-friendly and occasionally too time-consuming. Defining dropouts was a difficult task, since several studies defined this differently. We based our definition of adherence on the days the app was used, namely, at least 30% of the planned time. According to this definition, the dropout rate in the study (43%) is consistent with the work reported by Meyerowitz-Katz et al [38].

Accordingly, after analyzing the differences between the two groups, we observed that the participants’ TIRs improved regardless of whether they started with the intervention or the waiting phase. This may suggest that wearables alone are not sufficient to improve glucose management, but that extra feedback, in this case in the shape of an app, led participants toward better glucose management, although this cannot be determined definitively from the current data. The TIR of the start-group did not decrease significantly after they stopped using the PROTEIN app, suggesting there is a legacy effect for the users. The participants kept using the sensors as planned (12 weeks during intervention and 6 weeks in the waiting phase), which allowed us to compare the two periods of time. From these data, we observed patterns consistent with improved glucose management during the PROTEIN app use; however, causality cannot be inferred due to the exploratory design and potential selection bias. A dietary protocol before intervention could be added in further research to provide a better understanding of dietary change.

Furthermore, prior studies have shown that CGM is an important tool for the management of T2D [13,39,40]. When we analyzed our participants according to the assigned group, we can see that using wearables without guidance on how to effectively use them did not cause a change in the glucose levels. Interestingly, the participants who used the PROTEIN app right away (start-group) had a decrease in the TIR once they stopped using it, but this decrease was not statistically significant. This suggests that 12 weeks may not be enough to maintain a health-supporting behavior for a longer period. As no significant changes in weight were seen over the study period, improvements in TIR appear to be driven by the consumption of more adequate meals, independent of the group they were randomized into. Therefore, we cannot determine whether the PROTEIN app influences their caloric intake, but the findings are consistent with changes in food selection.

Overall, the PROTEIN application generally provided suitable meals for individuals with T2D and prediabetes. However, on a few occasions, the recommended protein amounts fell short when compared to the knowledge tracker. The first 2 weeks during which the participants were expected to train the app proved to be excessively time-consuming, leading to a decline in the compliance of our participants. Additional factors, including managing multiple devices and integrating the protocol into daily routines, were also identified as reasons for disengagement. The mobile app was programmed to change the meal plan for the whole week once the subject skipped a meal and added a new one instead. Moreover, the PROTEIN app sometimes failed to “learn” or adapt after two weeks, despite participants having the same breakfast daily.

The PROTEIN AI-advisor adjusted carbohydrate targets in response to low TIR and was programmed to exclude foods associated with repeated high excursions. We could not detect the frequency and magnitude of these adaptations in practice. Moreover, participants reported cases in which the system did not sufficiently adapt to repeated meal patterns, indicating that the current personalization was only partially realized and requires further optimization.

It is important to acknowledge that 21 participants had reliable CGM data, making the trial substantially underpowered and exploratory relative to the original sample size calculation with 300 participants. All suggestions of effect should be interpreted as exploratory signals rather than solid evidence.

Furthermore, limiting analyses to periods of active app use introduces substantial selection bias toward more engaged users, likely underestimates variability in real‑world use, and may overestimate potential effects.

Several limitations apply to our exploratory pilot study. Recruitment and retention were challenging, exacerbated by the COVID-19 pandemic and technical barriers (eg, Android-only compatibility). These factors limited the sample size and statistical power, highlighting the need for larger, more inclusive studies. Several prespecified outcomes were registered in the clinical trial registry but are not reported in this manuscript because of the limited sample size and incomplete data collection. A detailed comparison of registered and reported outcomes is provided in Table 2. Even though we saw a moderate difference in the TIRs while using the PROTEIN app, we were unable to assess participants’ TIR prior to using the wearables and/ or using the application, which could provide further robust evidence as to the health behavior changes.

Therefore, we cannot fully exclude the effects influencing intra-participant TIR patterns, as preintervention blind CGM data periods were not available.

Despite these limitations, the study provides preliminary indications of the feasibility and potential of combining mHealth, AI, and CGM data to support individualized diabetes care; nevertheless, it must be recognized that our own evidence indicates the intervention does not feasibly support increased TIR, except for a short period and for a small proportion of app users. This emphasizes important limitations regarding the broader clinical use and sustainability of the approach.

Our exploratory pilot study adds novel data on the feasibility, adherence patterns, and preliminary glycemic impact of such an AI-driven system under real-world conditions.

Future research should focus on improving user engagement, expanding compatibility, and evaluating long-term clinical outcomes in larger, more diverse populations.