Development of PI-ML Algorithm for Prediction of the Real-time Risk for Developing Pre-diabetes

Noncommunicable diseases (NCDs) such as cancer, cardiovascular diseases, and diabetes represent 74% of the disease burden globally and are the major causes of preventable premature deaths. (1,2).

Diabetes is a chronic NCD characterized by elevated glucose blood levels. In 2021, the prevalence of diabetes in Europe was 1 in 11 adults (61 million), a figure projected to rise to 69 million by 2045. (3) The global management and treatment of diabetes cost 988 billion dollars in 2021. Despite these expenditures, diabetes remains the third leading cause of death, accounting for 6.7 million deaths, with 1.1 millions of these in Europe alone. Among the different types, type-2 diabetes (T2D) comprises 90% of total diabetes cases, primarily presenting in adulthood. (3,4) Risk factors for T2D include genetic predisposition, family history, metabolic syndrome, obesity, physical inactivity, age, and ethnicity. There are 541 million adults worldwide with Impaired Glucose Tolerance (IGT), a significant risk factor for T2D. (5,6) IGT and Impaired Fasting Glucose (IFG) represent intermediate conditions within the "healthy-to-T2D transition" and are symptoms of prediabetes. (7,8) Notably, prediabetes represents an early-stage condition that can be reversed. Studies show that T2D progression can be reduced by approximately 58% within three years through lifestyle modifications. Physical activity of 30-54 minutes at least 2-5 days per week is recommended, as well as a healthy diet. (9) Efforts have been made to develop non-invasive diabetes risk prediction models based on clinically available parameters. (10) The onset of T2D involves complex, multiscale mechanisms starting from molecular, tissue, and organ levels, leading to dysfunction in physiological processes. Chronic inflammatory biomarkers play a significant role in T2D pathogenesis. Recognizing this multi-level approach is a step towards personalized disease diagnosis. However, there remain challenges related to modelling the "healthy-to-prediabetes transition" from both case study and methodological perspectives. (5,13,14) The objective of this project is the development of a prototype tool, aimed at real-time prediction of prediabetic risk. This tool will incorporate a series of patient-specific mathematical models simulating metabolism, pancreas hormone production, microbiome metabolites, inflammatory processes, and immune system response. These models were initially developed during the FP7 MISSION-T2D project and further developed into the implementation of an integrated, multilevel, and patient-specific model, incorporating genetic, metabolic, and nutritional data for the simulation and prediction of metabolic and inflammatory processes in the onset and progression of T2D. (14-18) The prediction algorithm will utilize a "physics-informed machine learning" (PIML) approach, combining a comprehensive dataset from both existing and new clinical trials, with continuous data input through wearable sensors. (19) The final algorithm will be hosted on a web-based platform where both medical professionals and patients can input data from multiple sources.

A dedicated prospective observational study described in this application will be conducted in Latvia recruiting adult participants with metabolic risk factors - overweight and obesity grade I, for data collection purposes to validate the developed machine learning PIML algorithm for pre-diabetes real-time risk prediction.

Data collections has three main purposes: I. Input Data for the in-silico MT2D model:

The input of the simulations includes the following discrete starting parameters: gender (M/F); weight; height; number of sessions of physical activity (0, 1, 2, 3); duration of the bout of physical activity (30, 60, 90 min); intensity in terms of % VO2max (40, 60); 3 meals per day; in each meal are specified the carbohydrates (low, medium, high), proteins (low, medium, high) and fats (low, medium, high).

II. Validation of the MT2D outputs: Output numerical values of the simulations from the model include:

1. Inflammation markers (recorded every 8h): B-cells (B-1, B-2), PBL, TH (Th1, Th2, Th17, Treg), CTL, Treg, NK, MA, DC, EP, ADIP (number, volume), IgM, IgG (IgG1, IgG2), IC, IL-2, IL-12, IFN-g, IL-4, TNF-a, TGF-b, IL-10, IL-6, IL-18, IL-23, IFN-b, IL-1b, LPS, leptin;
2. Metabolic outcomes (recorded every minute): arterial concentrations (glu, pyr, lac, ala, gly, FFA, tgl, O2, CO2); organs (22 metabolites); hormones (insulin, glucagon, epinephrine);fasting glucose; rate of appearance (glucose, alanine, triglycerides);total daily energy balance; %VO2max; anthropometric measures: BW, BMI, fat mass, fat free mass.

III. Data for training/validation of the physics-informed machine learning (PIML) algorithm: demographic Data; health-related Data; lifestyle Data (e.g., food consumption data and physical activity data); continuous ingestion through wearable sensors (Continuous Glucose Monitoring (CGM and tracker of physical activity e.g., Fitbit Charge 5, EDIBit.) These devices, such as smartwatches and fitness/activity trackers (e.g., Fitbit, Fibion, Apple Watch), are equipped with sensors that can track a variety of health metrics, including physical activity, heart rate, sleep patterns and increasingly also for glucose monitoring (non-invasive continuous glucose monitoring are still under development). (20,21) This data can be used to identify patterns and trends in a person's health, which can help with the early detection of diabetes/prediabetes. (22) Machine learning (ML) models show potential in enhancing early detection by analysing various risk factors and predicting outcomes. However, before these models are integrated into healthcare systems and clinical practices, they must be rigorously evaluated. One of the most robust methods for such evaluation is through external validation using longitudinal cohorts. (23-26).

Timeline and probands: The study will run for 18 months. During this period, 75 individuals will be followed for 4 months. The recruitment will take place from January 2024 until June 2025, parallel to the study period.

Subject identification: For each participant who has signed the Participant Informed Consent Form, the investigator must allocate a unique two letter and three-digit Participant Identification Number.

All documents, forms and data (including bio-materials - urine, blood and stool samples) files will be tagged with this Participant ID. Each participant, eligible and not eligible, will be documented in the Screening and Enrolment Log.

Data management: The data of the participants will be entered in an eCRFs. An eCRF will be provided by the project consortium partners Spindox Labs and CheckHealth. The eCRF will be maintained by staff at the University of Latvia. All study data will be captured in the eCRF and monitored by the monitor. All processes will be handled in accordance with standard operating procedures (SOPs).

Collected data from wearable sensors and remote-assisted questionnaires will be managed by the LinkWatch cloud-based storage for the entire duration of the project. This includes regular back-up procedures and security provisions in accordance with the GDPR and the CHK security policy. The security policy is based on the framework recommended by the Swedish Contingency Agency (Reg. 2016/679/UE), following ISO/IEEE 27000.

Paper forms of the questionnaires will be used due to possible technical problems of the platform. Consent form will be stored in archive at University of Latvia.

Macronutrient food intake will be obtained through a three-day food diary. That will be provided in different ways - initializing the macronutrient database with finish Finelli database, using Open food facts for packaged food. Meanwhile participants will be asked to double information, writing manually food diary for three days (two working days, one holiday).

In visit 1 and visit 3 food frequency questionnaire and in each visit, 24h recall will be obtained. Physical activity will be tracked using a commercially available FitBit Charge 5 tracker.

Between visit 1 and visit 2, specific recommendations will not be provided. This phase will serve as a data collection period to accurately assess participants' current lifestyle factors.

To enhance data diversity for PIML model training and validation, during the second visit, each participant will receive dietary and physical activity recommendations to modify lifestyle habits for the next two months.

Physical activity recommendations will follow a structured plan, ensuring a gradual increase in activity level personalized for each participant. The time period between the first and second study visits will serve as a data collection phase to accurately assess participants' current activity levels. Baseline physical activity levels, which will be used to establish personalized physical activity targets for each participant, will be calculated as the average minutes of moderate to vigorous physical activity (MVPA) recorded over the four weeks preceding the second visit. The initial weeks after the first visit will not be used for baseline activity calculation to limit possible behavioural changes participants might exhibit simply because they have started participating in the health related study.

These prescriptions will align with the relevant WHO guidelines and summarized in hand-out materials.

Blood samples from enrolled patients at the three cohort visits (Visit 1, Visit 2 and Visit 3) will be collected and initially processed to obtain plasma samples and temporarily stored (at -80°C) by the recruiting site (see below blood samples processing before shipment).

Plasma samples will be shipped to the project partner Italian Liver Foundation (FIF) in Basovizza, Trieste (Italy). At FIF, plasma samples will be stored at the at a temperature (-80°C) and access-controlled freezer for 10 years.

At FIF, batched samples will be unfrozen and further processed to determine plasma abundances of interleukins (IL-2, IL-6, IL-10, and leptin).

Blood will be collected for determination of clinical blood tests on the study site and for sample transfer to FIF for further biomarker analysis and preservation at the University of Latvia.

Participants will collect their faecal samples at home (together two samples - in first and second visit, respectively). The primary goal of faecal samples collection is to create a long-lasting PRAESIIDIUM biobank. In a subsequent moment, faecal samples might be processed for microbiome analysis. After arrival at local hospital sample should be frozen at -80°C until processing. Further processing - upon availability of additional funds (Fondazione Italiana Fegato ONLUS, Italy).