1386-P: Machine Learning Predicts T2DM Incidence Using Basic Clinical and Laboratory Parameters Free

Introduction and Objective: Artificial Intelligence (AI) and Machine Learning (ML) have the potential to improve risk prediction by identifying complex patterns in clinical and laboratory data, surpassing traditional approaches. ML has already shown success in detecting metabolic diseases, including Type 2 Diabetes Mellitus (T2DM). However, the ability to accurately forecast T2DM incidence is even more beneficial, enabling earlier interventions and treatment.

Methods: This observational study aimed to leverage ML to predict the 4-year risk of developing T2DM. A cohort of 904 cardiovascular risk patients, initially free of T2DM, was analyzed at baseline, with data including anthropometric measurements, clinical and laboratory parameters, and recent metabolic biomarkers. Over four years of follow-up, 10.2% of the patients developed T2DM.

Results: The ML approach, utilizing the Caret package in R, applied 50 variables. Patients were randomly split into training and test cohorts (75:25), with oversampling used to address class imbalance in T2DM incidence. Recursive feature elimination (RFE) was employed to identify the most relevant variables. A Support Vector Machine (SVM) model with a linear kernel demonstrated the most promising predictive performance, achieving a balanced accuracy of 73%, a sensitivity of 74%, a specificity of 71%, and an AUC of 0.727. The top-ranked predictors for T2DM were glucose measurements (2-hour OGTT glucose, fasting glucose, and HbA1c), HDL-cholesterol, and the triglyceride-glucose (TyG) index.

Conclusion: In conclusion, ML proves to be a valuable tool for identifying individuals at risk of T2DM, paving the way for personalized medicine through earlier diagnosis and tailored interventions.