Approach
The MELISSA project aims to fill the gap between Artificial Intelligence and its validated application in daily diabetes management routine by introducing the MELISSA platform - the world's first fully automated holistic AI-driven treatment personalisation and optimisation platform:
The ABBA algorithm and goFOOD – the foundations of MELISSA
Over recent years, key partners of the MELISSA consortium collaborated towards the development and in silico (meaning performed on computer or via computer simulation) validation of a novel AI-powered algorithmic approach for treatment optimisation and personalisation of insulin-treated people with type 1 diabetes, named Adaptive Basal-Bolus Advisor (ABBA).
They introduced an innovative algorithm that supports inputs from either Self-Monitoring Blood Glucose meters or Continuous Glucose Monitors to provide personalised suggestions for the daily Basal Rate (the rate at which an insulin pump infuses small doses of short-acting insulin) and three ICR per day (insulin-to-carbohydrate ratio - the amount of insulin needed for a fixed amount of carbohydrates to be processed) – one value for each of the three main meals – to people with type 1 diabetes either on insulin pump or Multiple Dose Insulin therapy.
To address the challenges related to the differences within and between patients and achieve personalisation of the insulin treatment, ABBA uses Reinforcement Learning. As a branch of Machine Learning, Reinforcement Learning is an intensively active research field which embraces algorithms able to learn from data and perform optimisation within uncertain environments. The field of Reinforcement Learning includes problems where an agent attempts to improve its performance over time at a given task through the continual interaction with its environment. ABBA employs the Actor-Critic method, a type of Reinforcement Learning, for updating Basal Rates and Insulin to Carb Ratios.
The Actor-Critic algorithm belongs to the class of Reinforcement Learning and is characterised by the separation of the agent into two complementary parts: the Critic, responsible for the control policy evaluation and the Actor, responsible for the control policy improvement. In Actor-Critic learning, the agent follows a certain control policy and performs transitions between states within an uncertain environment.
During MELISSA the ABBA will be further developed, extended and in silico validated to cover the needs of all insulin-treated people with diabetes on Multiple Dose Insulin therapy, including people with Type 2 diabetes. ABBA will be extended to also include additional lifestyle, conceptual and behavioural information as monitored by other external devices (smartphone, smartwatch) and quantified by a dedicated ‘in-house application’ (see below) and commercially available apps (e.g., Fitbit).
goFOOD™ is an app that was developed by researchers of the ARTORG Center of the University of Bern. It is capable of translating food images into nutrient information which can support diabetes self-management.
During the clinical validation of the MELISSA platform the goFOOD™ Lite will initially be used. The users can register a food or beverage (non-packaged) by capturing photos of a meal before and two after consumption. The user must place a specifically designed credit card-sized reference card near the plate or glass, which will be later used for the retrospective analysis by dietitians and the algorithm.
In case of a packaged product, the user has to simply take a photo of the barcode and choose the percentage of it that was consumed in 25% steps (0, 25, 50, 75 or 100%). The application then asks for the type of the meal, which is either a snack or a main meal (breakfast, lunch, dinner). The application also gives the opportunity to the user to view a history log, i.e., a list of all the items that were recorded.
The collected food or drink data will be retrospectively analysed by the goFOOD™ application. In addition to the food logging, the app uses AI and Machine Learning-related algorithms to predict the calories and macronutrients (carbohydrates, protein, fat) of each meal based on the images. First, automatic segmentation takes place. Then, the system proceeds with the food item recognition. The volume for each food item is then estimated by the system by using the captured photos. Finally, the calories and nutrients for each food item are retrieved from publicly available food databases and multiplied by the total volume of the items, to estimate the total calorie and nutrient of the meal which is outputted to the user.
The MELISSA platform with applications for people with diabetes and health care providers
The best performing algorithmic approaches of ABBA and prediction models will be integrated into the MELISSA platform which will support interfaces with the various available diabetes-treatment sensors (SMBG/CGM/FGM and pens), as well as with Fitbit (e.g., physical activity, HR, stress, SpO2) and goFOOD™ (food logging; calories, carbohydrate, fat and protein content estimation).
Furthermore, intelligent user interfaces will be used to allow end-users to interact with the MELISSA platform and provide additional information about their condition.
As a result, the MELISSA team will deliver two applications for personalised medicine to support both people with diabetes and health care providers to make informed decisions regarding treatment choices and adjustments:
For people with diabetes, a mobile application will be available in order to collect data, monitor the parameters related to their condition’s progress and to execute of the algorithms for treatment adjustment recommendations.
For health care providers, a web service will be available to collect patients’ data including lab and clinical information, to monitor the progress of the condition and to take decisions based on the models for prediction and risk assessment.