Heart Failure Phenotype Prediction using Engineered Features and Ensemble Learning

Abstract

A complex clinical syndrome, heart failure (HF) arises when the heart is unable to pump blood sufficiently to meet the body's metabolic and oxygen demands. This results in three main phenotypes: HFpEF (preserved ejection fraction), HFrEF (reduced ejection fraction), and HFmrEF (mildly reduced ejection fraction). The detection of abnormal ejection fraction is a crucial element for clinical treatment, and it can be efficiently addressed using advanced AI techniques. In this work, feature engineering is performed on heart dimensions (x, y), systolic and diastolic volumes, thereby leading to 8 cardiac groups, yielding 63 clinically relevant features from the EchoNet-Dynamic dataset. Further, ML and EL models, namely Random Forest (RF), XGBoost (XGB), CatBoost (CB), AdaBoost (Ada), Voting (Vot), Stacking, Extra Trees (ET), and Gradient Boosting (GBC), are utilized on these features, and their results are compared. Moreover, the CAMUS data set is also used to validate the models. Results reveal that the Voting ensemble model achieved the best performance, with an accuracy of 97.8% and an AUROC of 99.8% for multiclass classification.