The assessment of dietary advice delivered by a personalised mobile application to improve glucose control for adults with type 2 diabetes
Introduction
Finding an optimal diet that suits millions of people has proven to be challenging for nutrition research. Advances in technology, specifically AI, have enabled us to gather and process vast amounts of information as a tool to create a higher level of personalization and subsequently to boost diet adherence. While the number of nutrition apps is infinite, few provide adaptive evidence-based nutrition advice for patients with type 2 diabetes (T2D) residing in Germany. Therefore, as part of the PROTEIN project we aim to trial a dynamic personalized nutrition application to improve dietary compliance in T2D.
Objectives
We aim to customize meal plans according to data from the PROTEIN health expert team, the subjects’ preferences and their continuous glucose monitor (CGM) data. Our primary objective is to increase the time in range (TIR) of the study participants by 5%.

Methods
PROTEIN, a RCT, has a 3-month intervention period. Here, the subjects will use our application, a CGM and an activity tracker to collect essential data to provide personalization. The AI advisor is responsible for providing the meal plans and consists of a food and activity recommender system (FARS) and a reasoning-based decision support system (RDSS). The researchers on site will upload recent CGM data (7-14 days) that will be used to provide customized plans according to postprandial glucose excursions during the intervention period. If the TIR is below 70%, the goal „Decrease carbohydrates“ will be activated for next week’s recommendations. Foods, apart from vegetables and pulses, that caused high glucose levels (≥140 mg/dL over 2-4 hours) will activate a push notification directly to the user. If a food causes high glucose levels on two or more separate occasions, it will be excluded from the meal plans the following week.
Results
The app is being tested in virtual users to assure the appropriateness of the meal plans created by the AI advisor for each user. 300 subjects are planned to be tested.
Conclusions
The PROTEIN system creates adequate meal plans for virtual users with high glucose levels suggesting that an improvement of TIR can be achieved in real patients.
Funding
European Union’s Horizon 2020 research and innovation programme under grant agreement No 817732