Stanford University researchers have developed a novel approach to classify individuals with type 2 diabetes (T2D) into distinct metabolic subphenotypes using continuous glucose monitoring (CGM) data and machine learning. This method, detailed in Nature, could pave the way for more personalized and effective diabetes management strategies.
The study, part of the Precision Diets for Diabetes Prevention clinical trial (NCT03919877), enrolled 36 participants, with 32 completing all metabolic tests and included in the final analyses as the main study cohort. An independent cohort of 24 individuals completed OGTT and IST. The research team employed a combination of physiological tests, including frequently sampled oral glucose tolerance tests (OGTT), isoglycemic intravenous glucose infusions (IIGI), and insulin suppression tests, to quantify various aspects of glucose metabolism.
Identifying Metabolic Subphenotypes
The researchers focused on four key metabolic measures: muscle insulin resistance (IR), beta-cell dysfunction, incretin dysfunction, and hepatic IR. They calculated a standardized deviance score for each measure to identify an individual's most 'abnormal' physiological process. Participants were then classified into dominant metabolic subphenotypes based on these deviance scores.
Machine Learning for Subphenotype Prediction
To translate these findings into a more accessible diagnostic tool, the team developed a machine-learning framework to predict metabolic subphenotypes from OGTT glucose time series data. They extracted 14 features from the OGTT glucose time series, such as glucose level at time 0 (G0), 60 (G60), 120 (G120), 180 (G180), peak glucose level (G_Peak), length of the glucose time-series curve over the frequently sampled OGTT time interval (CurveSize), AUC, pAUC, negative (n)AUC, iAUC, coefficient of variation (CV), time from baseline to peak value (T_baseline2peak), slope between baseline to the peak glucose level (S_baseline2peak), and slope between glucose values at the peak and at the end (at t = 180 min) (S_peak2end).
The machine learning models were trained using data from the initial cohort and validated on an independent cohort. The models demonstrated high accuracy in predicting muscle IR, beta-cell dysfunction, incretin dysfunction, and hepatic IR. The performance of each model on each feature set and on each metabolic subphenotype was evaluated using auROC. We also calculated sensitivity, specificity, F1, precision and accuracy. Metrics were aggregated and summarized.
Clinical Implications
These findings suggest that CGM data, combined with machine learning, can provide valuable insights into the underlying metabolic abnormalities driving T2D. By identifying specific subphenotypes, clinicians may be able to tailor treatment strategies to address the unique needs of each patient.
"This approach could revolutionize diabetes care by moving away from a one-size-fits-all approach to a more personalized model," said a lead researcher on the study.
Future Directions
The researchers acknowledge that further validation in diverse populations is required to assess the generalizability of their findings. However, this study represents a significant step towards a more precise and individualized approach to diabetes management.
