Abstract 12897: Automated Assessment of Left Ventricular Function in Tetralogy of Fallot - A Machine Learning Approach

Abstract

Introduction: Congenital heart disease (CHD) is associated with significant cardiac anatomic and functional variability. This study tests the implementation and validation of a convolutional neural network algorithm (EchoNet-Dynamic) on a pediatric dataset via a transfer learning approach. We compare the echocardiographic systolic function measure, left ventricular ejection fraction (LVEF), with CMR as the gold standard. Methods: Patients ages 1-18 (44% female) who underwent a transthoracic echocardiographic evaluation and CMR for repaired Tetralogy of Fallot (TOF) were divided into 2 groups of EF < and >40%. 406 (2D-A4C views) videos with LVEF calculated using Bullet and Simpson method were identified. MRI EF as a ground truth label was compared with predicted EF data. Inference weights were derived from EchoNet-Dynamic and the weights from re-training the model on pediatric data. Correlation statistics were computed. The metrics included mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R2). Results: The R2 of actual vs predicted was 0.83 (0.71-0.89) with an MAE of 3.55. The predicted EF vs MRI EF resulted in an R2 of 0.48 and an MAE of 6.39 (Figure 1 a-c). Bland-Altman analysis between actual vs predicted EF estimates LVEF (bias = 1.64%) with 95% limits of agreement of -7.80% to 11.09%. MRI EF vs predicted values, the model estimates LVEF (bias = -3.12%) with 95% limits of agreement of -18.71% to 12.48%. The smaller body surface area and ages correlated with better model performance in both groups. Conclusions: EchoNet-Dynamic trained on a pediatric data set can accurately predict EF and this validates the model’s performance on a distinct CHD population. The predicted EF accurately compares with the MRI assessment. EchoNet Dynamic can be employed via a transfer learning method for automated real-time evaluations of cardiac function in CHD patients with compromised ventricular function and complex anatomical phenotypes.