Abstract P378: A Novel Risk Prediction Score for Incident Heart Failure Among Patients With Diabetes

Abstract

Introduction: Diabetes is a well-established risk factor for heart failure (HF). However, risk prediction models for incident HF in patients with diabetes are lacking. Accordingly, we aim to develop a novel risk-prediction model for HF in a diabetes cohort using a hybrid machine learning method for variable selection and CoxPH approach for risk prediction. Methods: We used data from the ACCORD trial that included patients with prevalent diabetes at baseline and had adjudicated HF events on follow up. The data was divided in to 80/20% training and testing (n = 6,896/1,725, respectively) and 152 baseline variables were included. A random survival forest (RSF) of 1000 trees was constructed to identify the top 20 predictors. Finally, a 5-year incident HF risk score was developed using the 9 variables selected by backward CoxPH regression. Results: The study included 8,621 participants (Mean age: 63 y, 62% male) without prevalent HF. The RSF model for incident HF with top 20 predictors derived from the training cohort had superior discrimination than the CoxPH model in the testing cohort ( C -index 0.80 vs. 0.78; P < 0.001). The top 9 variables identified included age, waist circumference, BP, serum creatinine, HDL, prior MI, prior CABG, and EKG QTC duration and T-axis. The weighted risk score based on the nine clinical variables was used to create a composite HF risk score ranging from -6 to 36. A 1-unit increment in the risk score was associated with a 27% higher risk of HF at 5 years. A graded increment in the 5-year HF incidence was noted across quintiles of the HF risk score ranging from 1.2% in the lowest to 9.8% in the highest quintile (Figure). The HF risk score demonstrated good discrimination with an overall C -index of 75.2. Conclusions: In conclusion, an RSF model in a clinical diabetes cohort outperformed conventional CoxPH in the identification of important risk factors of incident HF. Additionally, the diabetes-HF prediction model using common clinical variables demonstrated good discrimination and risk stratification in the overall cohort.