Abstract 16426: A Risk Model to Predict Atrial Fibrillation in Diabetes Using Machine Learning: The ACCORD Study

Abstract

Introduction: Atrial fibrillation (AF) is the most common cardiac arrhythmia seen in clinical practice, which necessitates identifying factors that can precipitate this dysrhythmia. We utilized the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial to build an AF risk prediction model specifically in the type 2 diabetes (T2DM) population, using machine learning (ML) approaches, and comparing to the established CHARGE-AF risk model. Methods: We utilized a patient cohort from the ACCORD clinical trial without prior AF. Baseline characteristics were collected as potential risk factors and incident AF was defined from ECG at follow-up. A random forest (RF) binary classifier was trained with 27 variables, including age, race, sex, smoking, waist circumference, BMI, blood pressure (BP), history of heart disease, medications, cholesterol, renal function, and left ventricular hypertrophy by ECG. All ACCORD subjects had T2DM, therefore this variable was not included. The CHARGE-AF model consisted of a Cox proportional hazard model trained on the following variables: age, race, height, weight, systolic/diastolic BP, current smoking, heart failure, prior MI, and BP medication. Area under ROC curve (AUC) was measured with randomized, stratified, 5-fold cross validation with 80/20 splits. Imputation and resampling methods were used for missing data and class imbalance. Results: Out of 9307 patients in the ACCORD cohort, 175 developed AF during the study period of 6.26 years, with a mean time to AF event of 3.68 years. The RF model, with an AUC of 0.912, showed improved prediction of AF over the Cox model, which had an AUC of 0.784. Permutation feature importance revealed age, race, waist circumference, weekly alcohol consumption, sex, and GFR as the most predictive features of the RF model (Fig 1). Conclusions: The CHARGE-AF risk model performed modestly on the ACCORD cohort, while a RF model showed superior performance.