Deep Learning Predicts Heart Failure With Preserved, Mid-Range, and Reduced Left Ventricular Ejection Fraction From Patient Clinical Profiles

Background: Left ventricular ejection fraction (LVEF) is the gold standard for evaluating heart failure (HF) in coronary artery disease (CAD) patients. It is an essential metric in categorizing HF patients as preserved (HFpEF), mid-range (HFmEF), and reduced (HFrEF) ejection fraction. Objectives: We sought to investigate the effectiveness of using deep learning as an automated tool to predict LVEF from patient clinical profiles using regression and classification trained models. Methods: Data from 303 CAD patients were obtained from American and Greek patient databases and categorized based on the American Society of Echocardiography and the European Association of Cardiovascular Imaging (ASE/EACVI) guidelines into HFpEF (EF > 55%), HFmEF (50% ≥ EF ≥ 55%), and HFrEF (EF > 40%). Clinical profiles included 13 demographical and clinical markers grouped as cardiovascular risk factors, medication, and history. Two deep learning-based models were then developed and trained using convolutional neural networks (CNN) to estimate LVEF levels from the clinical information and for classification into one of three LVEF-based HF categories. Results: A total of seven clinical markers were found important for discriminating between the three HF categories. Using statistical analysis, diabetes, diuretics medication, and prior myocardial infarction were found statistically significant (p-value < 0.001). Furthermore, age, body mass index (BMI), anti-arrhythmics medication, and previous ventricular tachycardia were found important after projections on the ArcViz convex hull with an average nearest centroid (NC) accuracy of 94%. The regression model estimated LVEF levels successfully with an overall accuracy of 90%, average root mean square error (RMSE) of 4.13, and correlation coefficient of 0.85. A significant improvement was then obtained with the classification model, which predicted HF categories with an accuracy ≥ 93%, sensitivity ≥ 89%, 1-specificity < 5%, and average area under the receiver operating characteristics curve (AUROC) of 0.98. Conclusions: Our study paves the way towards implementing deep learning-based models clinically to ensure faster, yet accurate, automatic prediction of HF based on LVEF with only clinical profiles and corresponding information as input to the models. Invasive, expensive, and time-consuming clinical testing can thus be avoided, enabling reduced stress in patients and simpler triage for further intervention.