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

UNLABELLED: An inflection point (IP) marking accelerated β-cell decline occurs ∼1-2 years before type 1 diabetes diagnosis. Precisely determining this timing could optimize clinical intervention. We developed machine learning models to predict proximity to the inferred metabolic IP using oral glucose tolerance test (OGTT) data from islet autoantibody-positive individuals in the TrialNet Pathway to Prevention study. OGTTs were retrospectively labeled by estimating time to the IP using thresholds from 1.2 to 1.6 years. Models were trained on engineered glucose and C-peptide dynamics, slopes, and composite indices, with feature selection by recursive feature elimination (RFE). Classifiers included support vector machines (SVM), random forests, and gradient boosting; a Cox proportional hazards model was also applied to estimate visit-level time to diagnosis and derive time-to-IP predictions. External validation was performed in the independent Diabetes Prevention Trial-Type 1 cohort. The best-performing model-SVM with RFE at the 1.4-year threshold-achieved an area under the curve of 0.77 (95% CI 0.72-0.82). The Cox model provided complementary numeric estimates of time to diagnosis and time to IP, sensitive to the cohort-level lead time offset (Δ). These findings show that machine learning and survival analysis can support early metabolic shift detection, enabling timely risk stratification and personalized monitoring in at-risk individuals.

ARTICLE HIGHLIGHTS: 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.