Predictive Modeling for Diabetes Subtype Classification in India: A Machine Learning Approach

Abstract Objectives: We aimed to develop a predictive model to effectively distinguish between type 1 (T1D) and type 2 diabetes (T2D) in adolescents and young adult populations aged 10–30 years. Materials and Methods: Diabetic individuals aged 10–30 years of Asian Indian descent ( n = 10,325) were identified from electronic medical records spanning January 1992 to March 2021. T1D was defined by a history of diabetic ketoacidosis, C-peptide levels <0.6 pm ol/mL, and necessity for insulin uptake from the time of diagnosis. Conversely, T2D was distinguished by the absence of ketosis, adequate beta-cell reserves evidenced by C-peptide levels (>0.6 pm ol/mL), and a positive response to oral hypoglycemic agents for more than 2 years. Utilizing logistic regression, we developed a multivariable classification model incorporating clinical parameters (age at diagnosis, body mass index [BMI], and family history of diabetes), biochemical parameters (lipid profile and hemoglobin A1C [HbA1c]), and biomarkers (glutamic acid decarboxylase [GAD] antibody status) to differentiate between T1D and T2D. Results: In the development cohort, 22% of participants had T1D. Each predictor feature effectively distinguished between T1D and T2D. Model performance was high, with the area under the receiver operating characteristic curve (ROC AUC) ranging from 0.92 (using only BMI and age at diagnosis) to 0.97 (utilizing all predictors). Validation results remained consistent, with ROC AUC ranging from 0.92 to 0.96. Conclusion: The integration of clinical characteristics and GAD antibody status into the classification model yields high accuracy in differentiating between T1D and T2D in Asian Indians. This model can serve as a valuable tool to aid physicians in effectively classifying diabetes subtypes and planning treatments in South Asians.