1357-P: Machine-Learning Models to Guide Preventive Interventions for Type 2 Diabetes in India

Introduction and Objective: The rising global prevalence of type 2 diabetes (T2DM) is linked to obesity. However, in India, fewer than half of adults with T2DM are obese. We used machine learning (ML) to identify novel risk factors and targets for prevention in an Indian population. Methods: We used multiple ML models (i.e. Logistic Regression, Decision Tree, XGBoost) to study T2DM using data on 7067 participants and 997 variables from the cross-section Indian Migration Study, conducted in 4 regions of India from 2005-2007. T2DM was defined per ADA. The classification metrics (F1 and AUC) were chosen as primary optimization metrics due to the imbalanced nature of the diabetes classes. Key risk factors were identified through coefficient scores and feature importance analyses, and their effects were visualized using SHAP dependency plots. Results: The most accurate model achieved an F1 score of 0.60 and an AUC of 0.885 for T2DM. Prior to the addition of genetic (SNP) data, F1 score was 0.48. Top features associated with T2DM were: regular medication use, prior tuberculosis, greater intake of glutamic acid, arginine and leucine, and lower intake of carbohydrates and protein. Having the rs10811661_TT SNP was also associated with T2DM. Thyroid disease and living in eastern India were inversely associated with T2DM. Conclusion: High-dimensional ML models trained on epidemiological data may help guide preventive interventions in India and identify individuals at high risk for T2DM. N.A. Mankodi: None. A. Tewari: None. E. Gitlin: None. A. Aggarwal: None. P. Chebrolu: None. Mastercard Research Assistance for Primary Parents (K23DK136388)