Abstract P131: Individualized Risk Prediction for Type 2 Diabetes: A Secondary Analysis of the Diabetes Prevention Program

Abstract

Introduction: Type 2 diabetes (T2D) risk prediction models that predict risk with individualized preventive intervention effects can help targeted prevention. Hypothesis: An individualized T2D risk prediction model developed in the Diabetes Prevention Program (DPP) randomized trial that incorporates differential treatment effects based on an individual’s fasting glucose (FG) level and body mass index (BMI) will provide more accurate predictions than a model without individualized treatment effects. Methods: We included 2640 DPP participants randomized to the placebo, metformin, or lifestyle arms. Using 50% of the DPP sample, we developed a non-individualized Cox model predicting T2D risk over 3 years with adjustment for sex, hemoglobin A1c, triglycerides, FG, BMI, and treatment, and an individualized model with adjustment for these variables plus interactions between treatment and age, FG, and BMI. In the remaining 50%, we evaluated the prediction accuracy of both models using the concordance (C)-statistic. We repeated this process 100 times with different random splits. We externally validated models among Multi-Ethnic Study of Atherosclerosis (MESA) adults with prediabetes (n=1067) where model predictions were computed for each intervention scenario and averaged for all participants. Results: Mean (standard deviation) age of participants was 51 (11) years in the DPP and 64 (10) in MESA, with 67% and 54% women, respectively. Mean C-statistics for our individualized and non-individualized prediction models (100 splits) in DPP were 0.71 and 0.70, respectively (table). The individualized model predicted higher benefit of lifestyle intervention at lower BMI levels and higher benefit of both metformin and lifestyle at higher FG. The C-statistic of the individualized model was 0.78 in the MESA validation. Conclusion: Creating individualized estimates of intervention effects can improve accuracy of T2D risk prediction models, allowing patients to make more informed decisions about prevention strategies.