Abstract MP58: Atherosclerotic Cardiovascular Events Prediction Using Machine Learning Models: Results From Action to Control Cardiovascular Risk in Diabetes Trial

Abstract

Listen

Translate article

Background: The AHA/ACC published a pooled cohort 10-year atherosclerotic cardiovascular disease (ASCVD) risk calculator to estimate the probability of initial ASCVD events based on statistical modeling. Our current study aimed to develop machine-learning models with the identical predictors among T2DM patients from the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial and further compared their performance in predicting composite outcome of myocardial infarction, non-fatal stroke and cardiovascular death. Methods: The guideline risk calculator provided 9 predictors including baseline age, gender, race, SBP, antihypertensive medication use, total cholesterol, high-density lipoprotein cholesterol, current smoking status and diabetes mellitus status. We developed three ML ASCVD Risk Calculators based on Linear Model (LM), Supporting Vector Machine (SVM) and Random Forest (RF) algorithms using 10-year follow-up data from ACCORD with the same 9 predictors. T2DM patients with prior ASCVD or with invalid follow-up time were excluded in our analysis. Those who had not experienced the composite outcome by the end of year 10 would be labeled as censored. 5-fold stratified random split was applied as a cross-validation strategy. Results: A total of 6581 T2DM participants were included in our final sample, with a mean age of 62.9±5.9 years old (range 51-79 yr), 44.1% female and 60.8% white. Among those, 12.2% (n=802) had developed composite ASCVD during a median follow up of 9.1 years. The performance AHA/ACC 10-year Risk Calculator was modest with AUC=0.604. In contrast, ML models showed better performance from validation data with LM AUC=0.854, SVM AUC=0.848, and RF AUC=0.866 (Figure). Conclusion: The ML ASCVD Risk Calculator outperforms the AHA/ACC pooled 10-year ASCVD risk calculator in predicting the composite ASCVD outcomes among those with DM from ACCORD trial. Future studies need to validate ML algorithms in other cohorts and further explore other potential valuable predictors.