Abstract
BACKGROUND: About one-third of patients with heart failure (HF) with reduced ejection fraction (HFrEF) may demonstrate left ventricular ejection fraction (LVEF) recovery with medical management. In this study, we developed a machine-learning (ML) model to predict LVEF improvement and compared its performance to a logistic regression (LR) model in internal and external validation cohorts. METHODS: We identified 3124 patients with HFrEF and ≥2 echocardiograms taken ≥6 months apart. Patients were split into development (n = 1812) and external testing (n = 1312) cohorts by site. The ML and LR models were trained using 49 features, with internal 5-fold cross-validation. Prediction performance and calibration in the internal and external testing cohorts were assessed using the area under the curve (AUC) and Brier score, respectively. RESULTS: LVEF recovery defined as an absolute LVEF increase of ≥10% occurred in 36.0% of the development cohort and 39.8% of the external testing cohort. The ML model performed better in predicting LVEF recovery compared to the LR model in the development cohort (AUC 0.719 vs 0.700, p = 0.045), however, there was no significant difference in external testing (AUC 0.702 vs 0.696, p = 0.498). Lower baseline LVEF and left ventricular dimensions, younger age and non-ischemic etiology contributed the most to the prediction of LVEF improvement. CONCLUSION: A ML model using readily available clinical and echocardiographic data did not perform better than traditional LR in predicting LVEF improvement when tested externally. Larger studies, including additional variables or alternative approaches, may improve LVEF recovery prediction.