Predicting the Timing of the Metabolic Inflection Point in Type 1 Diabetes Progression Using Machine Learning and Survival Analysis Models

We undertook this study to improve early identification of the metabolic inflection point (IP) preceding clinical type 1 diabetes in autoantibody-positive individuals. We aimed to develop and validate machine learning models using oral glucose tolerance test-derived dynamic features to detect proximity to the IP. A support vector machine trained on TrialNet Pathway to Prevention and tested on Diabetes Prevention Trial-Type 1 achieved an area under the curve of 0.77 at 1.4 years prior to diagnosis, with strong calibration and interpretability. Additionally, a Cox proportional hazards model provided numeric estimates of time to IP, offering complementary predictions. These results can support earlier intervention and timely monitoring through personalized oral glucose tolerance test-based risk stratification.