1288-P: Predictive Modeling of Type 2 Diabetes and Complications Using Machine Learning

Abstract

Background: Given the prevalence of T2D worldwide, early detection and prevention of both T2D and its complications are of utmost importance. We propose comprehensive person-centered machine learning (ML) risk prediction models for T2D and complications. Methods: From the UK biobank (UKBB), we identified a prevalent T2D group at baseline (9,136 cases and 108,615 controls), an incident T2D group over 10 years of follow-up (29,845 cases and 354,818 controls), and an incident diabetic kidney disease (DKD) group (4435 cases and 33083 controls). We included 3 types of risk factors: clinical (biomarkers and diseases with known diabetes risk like CVD), genetic in the form of polygenic risk scores (PRS) for the pathogenesis of T2D and other clinical risks, and lifestyle. For each group, we trained a model using decision-tree-based ensemble ML algorithm. We then used SHAP values to determine feature importance and chose the top 10 predictors to build a final model. Results: The T2D prevalence model included 74 features and had an AUC of 0.97, and area under the precision-recall curve (PR AUC) of 0.83. The final model with the top features (HbA1c, T2D PRS, primary HTN, LDL, major dietary changes in the last 5 years, age, plasma glucose, waist circumference, and total cholesterol) had an AUC of 0.98, and PR AUC of 0.88. The T2D incidence model included the same 74 features and had an AUC of 0.92, and PR AUC of 0.57. The final model with the top features had an AUC of 0.93, and PR AUC of 0.65. The DKD risk prediction model comprised 85 features, and it had an AUC of 0.81, and PR AUC of 0.37. The final model with the top features (Cystatin C, primary HTN, eGFR, age, HbA1c, T2D PRS, urea, BMI, BMI PRS, NAFLD PRS) had an AUC of 0.80, and PR AUC of 0.39. Conclusions: This study is an ongoing work; our interim results show the capability of ML to identify relevant risk factors and that comprehensive ML models could help clinicians decide on more precise and individualized T2D screening and treatment approaches to prevent T2D and complications. Disclosure A.Khattab: None. S.Chen: None. H.J.Sadaei: None. A.Torkamani: None.