New method promises personalized diagnostics and more effective diabetes treatment

New Method Promises Personalized Diagnostics and More Effective Diabetes Treatment

Type 2 diabetes is a heterogeneous disease whose pathophysiology is caused by various disturbances in glucose metabolism. However, current diagnostic methods, according to some experts, do not provide sufficient information about individual metabolic deviations. A research team led by Ahmed A. Metwally from Stanford University proposes a method, published in Nature Biomedical Engineering, that identifies metabolic subphenotypes of type 2 diabetes using continuous glucose monitoring and machine learning algorithms. The new method could significantly enhance personalized diagnostics and more effective therapy.

More than 537 million adults worldwide suffer from type 2 diabetes. However, the current study highlights that prediabetes, a state characterized by blood sugar levels higher than normal but not yet high enough to diagnose type 2 diabetes, is equally important. Although both conditions share a similar basis, current classification does not reflect individual variability in glucose dysregulation pathophysiology among patients.

New Perspective and Technology

Until now, scientists have usually assumed that type 2 diabetes is equally caused by insulin resistance and dysfunction of β-cells in the pancreatic islets, responsible for insulin production. However, in the current study, the research team hypothesizes that individuals with prediabetes and early stages of type 2 diabetes exhibit varying degrees of insulin resistance and β-cell dysfunction, as well as different defects in incretin action and hepatic glucose regulation.

They therefore propose classifying patients based on their underlying metabolic physiology rather than solely on current glycemic values These physiological foundations are present even before the onset of hyperglycemia but are difficult to identify outside specialized centers According to the research team, it is therefore necessary to develop a cost-effective method capable of identifying these individual factors

The oral glucose tolerance test (OGTT) is currently the most common method for measuring glycemic disturbances, standardized and globally used for over 100 years. Because OGTT can measure various metrics, such as glycemia at specific time intervals and curves with multiple time points, this method has significant potential for assessing the metabolic pathophysiology of individual patients.

However, analyzing multiple time points in OGTT tests is quite demanding In their study, scientists therefore combined OGTT data with continuous glucose monitoring (CGM) commonly used in home settings, evaluating these data using artificial intelligence and machine learning tools

Research Process

The researchers included 56 individuals without a prior history of diabetes, with fasting blood glucose levels below 126 mg/dl (7 mmol/l). Of these, 33 were classified as normoglycemic, 21 with prediabetes, and 2 with type 2 diabetes. The researchers subsequently created 3 cohorts: one for model training and testing, one validation cohort, and a home-testing CGM cohort comprised of participants from both the first and second cohorts.

To characterize dynamic glucose profiles during OGTT, glucose concentrations were measured at 5–15-minute intervals (16 time points) over 180 minutes after administering 75 g of oral glucose load. Subsequently, glucose curves from the 16-point OGTT performed in a hospital setting and the average of two home-based OGTTs using standard CGM devices were evaluated.

Then they conducted extensive metabolic profiling to assess the significance of four primary physiological phenotypes of glucose regulation disorders: insulin resistance (IR) in muscles, β-cell dysfunction, impaired incretin action, and hepatic IR. This allowed them to develop a machine learning algorithm utilizing glucose time-series data to predict these phenotypes.

Main Findings

In 32 individuals from the initial cohort, muscle or hepatic IR phenotypes were identified in 34%, while β-cell dysfunction or impaired incretin activity was found in 40%

Machine learning models created from OGTT data of these 32 individuals could predict glucose regulation disorder phenotypes with an area under the curve (AUC) of 95% for muscle IR, 89% for β-cell deficiency, and 88% for impaired incretin action. When conducting home-based OGTT using CGM (29 individuals), the models predicted muscle IR phenotype with an AUC of 88% and β-cell deficiency phenotype with an AUC of 84%

Early Identification of At-Risk Individuals

The research team demonstrated that metabolic physiology underlying glycemic disturbances varies significantly among individuals, and dominant metabolic subphenotypes of type 2 diabetes exist. Identifying these individually distinct metabolic subphenotypes using OGTT and widely available continuous glucose monitoring can improve the early identification of at-risk individuals who could then undergo targeted therapy and lifestyle adjustments to prevent the development of type 2 diabetes.

However, according to the study authors, further studies are needed to determine whether this method can also be applied to individuals with more advanced hyperglycemia and type 2 diabetes

Editorial Team, Medscope.pro

Source: Metwally A. A., Perelman D., Park H. et al. Prediction of metabolic subphenotypes of type 2 diabetes via continuous glucose monitoring and machine learning. Nat. Biomed. Eng (2024), doi: 10.1038/s41551-024-01311-6.