Mobile Artificial Intelligence Solution for Diabetes Adaptive Care

Project Information

MELISSA

Grant agreement ID: 101057730

DOI

10.3030/101057730

EC signature date 23 May 2022

Start date 1 June 2022

End date 31 May 2027

Funded under

• Health

Total cost

€ 6 061 709,00

EU contribution

€ 5 986 709,00

Coordinated by

UNIVERSITEIT MAASTRICHT
Netherlands

Periodic Reporting for period 2 - MELISSA (Mobile Artificial Intelligence Solution for Diabetes Adaptive Care)

Reporting period: 2023-12-01 to 2025-05-31

Summary of the context and overall objectives of the project

Despite advances in insulin delivery and glucose monitoring technology, achieving near-normal glycaemic control remains highly challenging for most people with type 1 or type 2 diabetes requiring intensive insulin treatment. Daily insulin needs for people with diabetes (PwD) fluctuate due to factors like carbohydrate (CHO) intake, physical activity, health conditions, and variables like mood, temperature, and insulin sensitivity. These factors contribute to hyper- and hypoglycaemic excursions from ‘erroneous’ insulin dosing, hindering optimal glycaemic control and quality of life. Consequently, many PwD do not reach recommended glycaemic targets and remain at increased risk of developing devastating short- and long-term complications related to ‘poor’ glycaemic control and hypoglycaemia. At present, systems for decision support with regards to daily insulin dosing for PwD treated with Multiple Daily Injections (MDI) are limited to the coverage of basal insulin requirements and to simple bolus calculators of meal-related insulin administration working with fixed algorithms that rely heavily on self-estimated CHO intake. Artificial Intelligence (AI)-enabled algorithms may have a considerable potential to further improve daily decision-making for many PwD by compensating for the effect of both known and unknown factors affecting insulin needs that the person may not be able to manage. MELISSA’s main objective is to provide a clinically validated, efficient and cost-effective digital diabetes management solution by combining AI-based approaches and the quantification of lifestyle and behavioural factors. The goal is to support PwD on insulin in providing personalized treatment and care recommendations for achieving and maintaining normoglycaemia.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

MELISSA is structured into work packages (WPs) focused on different project aspects, and established a Patient Advisory- (PAC) and Communication Committee (CC). PAC meets regularly and provides valuable feedback to the MELISSA researchers and CC focusses on the wider public outreach. The AI-powered Adaptive Basal–Bolus Advisor (ABBA) - a self-learning system based on actor-critic, model-free reinforcement learning (RL) - was further optimised by leveraging both current and historical glucose values and incorporating insulin-on-board accounting, and extended to support PwD on MDI using SMBG. The updated ABBA was validated in silico using an FDA-accepted virtual population of people with type 1 and type 2 diabetes, showing ~5% increase in time spent in glycaemic range and reductions in times spent above and below range compared to standard bolus calculators. The goFOODTM food segmentation, recognition, and 3D reconstruction features were optimized, and goFOODTM was localized for each trial country’s eating habits. The system now consists of two forms, i.e. goFOODTMLite (only data collection) and the full version of goFOODTM (data collection & quantitative estimation of CHO). The MELISSA App integrating goFOODTM and ABBA designated for the clinical trial has been released on the Google Play store. Multilingual User Requirements Questionnaires for PwD and healthcare professionals have been developed and are being distributed, including at institutions requiring ethical approval for the PwD version. The clinical validation study design for MELISSA has been finalized, with the protocol written and submitted for ethics approval across all study sites. Ethics approval is pending at all sites and partial ethics approval has been granted in Denmark and Germany. Standard Operational Procedures have been finalised, the Data Management Plan completed and the data management system has been built in the RedCap platform. Contracts with manufacturers like Novo Nordisk (for NovoPen 6 and Echo Plus smartpens and provision of insulin), Dexcom (for G7 sensors and receivers), and Withings (activity tracker) have been made for communication with the MELISSA app. Several manuscripts have been published in scientific journals and abstracts have been presented at scientific conferences. The MELISSA solution has been analysed using the business canvas framework, establishing its value proposition and ensure alignment with the true needs of users. In addition, two ethics reports for providing feedback on the project have been delivered. In line with MELISSA’s fifth objective, a sub-study on (digital) health literacy and numeracy is ongoing in the Netherlands, Denmark, and Greece, while in Germany ethical approval is still pending. Finally, a feasibility study has been completed in Switzerland to demonstrate the feasibility of using ABBA in daily practice by PwD.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

ABBA (Adaptive Basal–Bolus Advisor): Initially developed for PwD using CGM and insulin pumps, ABBA was subsequently adapted for SMBG with pumps and pens and is now being extended to all MDI-treated individuals. ABBA employs an actor–critic, model-free RL approach tailored to MDI therapy, addressing key challenges in RL, so it converges faster and provides consistent dose recommendations in real-world–like settings. This innovation is covered by a patent application.
goFOOD™: goFOOD™ leverages AI, computer vision and smartphone sensors (Android and iOS) to deliver automated food segmentation, recognition, and 3D reconstruction, enabling estimation of calories and macronutrients comparable to expert assessments, and has been localised to country-specific diets.
Integrated impact (ABBA + goFOOD™): Coupling automatic meal quantification (goFOOD™) with adaptive insulin dosing (ABBA) enables unprecedented personalisation of insulin adjustment for MDI therapy, closing the loop from meal detection → CHO estimation → dose recommendation, and represents a substantial advance in personalised diabetes care.

melissa-2nd-reporting-image-final.jpg