Machine learning to predict in-hospital outcomes in patients with acute heart failure

Abstract Background Acute heart failure (AHF) is a leading cause of mortality worldwide and a major public health issue with a still high rate of in-hospital outcomes. Physicians need more investigations and new tools to prevent those high-risk patients from major adverse events (MAE). While few scores are available for risk stratification of patients hospitalized for AHF using traditional statistical methods, the potential benefit of machine-learning (ML) is not established. Purpose To investigate the feasibility and accuracy of a machine-learning (ML) model using clinical, biological and echocardiographic data to predict in-hospital MAE in patients hospitalized for AHF and compare its performance with traditional models and existing scores. Methods The study cohort consists of consecutive patients admitted for AHF included in the French nationwide, multicenter, prospective, ADDICTO-USIC study involving 39 centers from 7 to 22 April 2021. Traditional clinical, biological, electrocardiographic and echographic data as well as a standardized exhaled carbon monoxide (CO) measurement and the presence of illicit drugs determined through an urine drug assay were recorded. Three ML models were developed using clinical and echocardiographic parameters to predict in-hospital MAE, including death, resuscitated cardiac arrest or cardiogenic shock requiring medical or mechanical hemodynamic support. Least absolute shrinkage and selection operator (LASSO) regression was used to select variables and prevent model overfitting. The ML models (LASSO, random forest and XGBoost) were then trained on 70% of patients and evaluated on the other 30% as internal validation. Their performance was compared against standard logistic regression model, using receiver operating characteristics (ROC) and precision-recall (PR) curves and area-under-the curves (AUC). Results Among 459 consecutive patients included (age 68±14 years, 68% male), 47 had in-hospital MAE (9.8%). Out of 28 clinical, biological, ECG, and echocardiographic variables, seven were selected as being the most important in predicting MAE in the training set (N=322): mean arterial pressure (MAP), ischemic cardiomyopathy etiology, sub-aortic velocity time integral (VTI), E/e’, tricuspid annular plane systolic excursion (TAPSE), illicit drugs and Carbon monoxide. The random forest model showed the best performance compared with the other ML models (AUROC=0.82, PR-AUC=0.48, F1 score=0.56). Our ML-score exhibited a higher AUC compared with an existing score for prediction of MAE (AUROC for ML score: 0.82 vs Acute HF-score: all p<0.001). Conclusions The random forest ML-model including seven clinical and echocardiographic variables, including carbon monoxide level and illicit drugs use, exhibited a better performance than traditional statistical methods or existing scores to predict in-hospital outcomes in patients admitted for AHF.Feature selection by LASSOPerformances of ML models on test set