(B47) Cardiovascular risk stratification in patients with type 2 diabetes

Publication Number: 657

Background: Newer glucose-lowering drugs have shown cardiovascular (CV) benefits in patients at high risk of CV disease (CVD). Currently, there is no CV risk stratification score specifically developed in patients with type 2 diabetes (T2D) that allows risk stratification using electronic healthcare data in the US.

Objectives: To develop and validate a CV risk stratification model to improve confounding control and study effect modification by CV risk in pharmacoepidemiologic studies.

Methods: Using the Medicare Research Practice Data Registry (RPDR) 2013-17 [claims linked to electronic health records (EHR)], we identified patients with T2D aged >65 years. Cohort entry was the date of a physician visit (a random date if a patient had >1 eligible visits). We identified CV outcomes (composite of hospitalization for heart failure, myocardial infarction, stroke, unstable angina, or coronary revascularization) within 1 year of cohort entry, using validated claims-based algorithms. Separately in patients with and without baseline CVD, we developed and compared outcome predictive models including claims-based variables only (e.g., comorbidities, medications, frailty score, measures of healthcare system use) vs. claims plus EHR variables (e.g., vitals and labs data). We trained models using data from one healthcare system and tested the model using data from another system, within RPDR. After cross-validation for model tuning, we used least absolute shrinkage and selection operator (LASSO), and extreme gradient boosted (XG-boost) models. Diagnostics included c-statistics to assess discrimination and brier score to assess calibration (lower scores are better).

Results: Among patients with CVD, we identified 13,613 and 12,422 patients in the training and testing sets (mean age, 77 years). XG-boost claims-only models had a slightly better performance than LASSO claims-only models, with c-statistics of 0.74 and Brier score of 0.01. The most influential predictors were baseline heart failure, loop diuretics and number of CV hospitalizations. In patients with no baseline CVD, we identified 10,224 and 8,867 patients in the training and testing sets (mean age, 74 and 73 years). XG-boost claims-only models performed slightly better than LASSO models, with c-statistics of 0.71 and Brier score of 0.01. The most influential predictors were age, combined comorbidity, and frailty scores. Claims-EHR models had similar performance as claims-only models in both patients with and without CVD.

Conclusions: In older T2D patients, CV predictive models showed good discrimination and accuracy, with slightly better performance in patients with baseline CVD. Adding EHR variables did not improve the model performance.