Machine Learning for Pediatric Echocardiographic Mitral Regurgitation Detection
Fei Feng,Mehreen Iqbal,Yong Fu,Amy Sanyahumbi,Shiying Hao,Craig Sable,Jiajia Luo,Lindsay A. Edwards,Doff B. McElhinney,X. Bruce Ling
DOI: https://doi.org/10.1016/j.echo.2022.09.017
IF: 7.722
2023-01-01
Journal of the American Society of Echocardiography
Abstract:BACKGROUND: Echocardiography-based screening for valvular disease in at-risk asymptomatic children can result in early diagnosis. These screening programs, however, are resource intensive and may not be feasible in many resource-limited settings. Automated echocardiographic diagnosis may enable more widespread echocardiographic screening, early diagnosis, and improved outcomes. In this feasibility study, the authors sought to build a machine learning model capable of identifying mitral regurgitation (MR) on echocardiography.METHODS: Echocardiograms were labeled by clip for view and by frame for the presence of MR. The labeled data were used to build two convolutional neural networks to perform the stepwise tasks of classifying the clips (1) by view and (2) by the presence of any MR, including physiologic, in parasternal long-axis color Doppler views. The view classification model was developed using 66,330 frames, and model performance was evaluated using a hold-out testing data set with 45 echocardiograms (11,730 frames). The MR detection model was developed using 938 frames, and model performance was evaluated using a hold-out testing data set with 42 echocardiograms (182 frames). Metrics to evaluate model performance included accuracy, precision, recall, F1 score (average of precision and recall, ranging from 0 to 1, with 1 suggesting perfect precision and recall), and receiver operating characteristic analysis.RESULTS: For the parasternal long-axis view with color Doppler, the view classification convolutional neural network achieved an F1 score of 0.97. The MR detection convolutional neural network achieved testing accuracy of 0.86 and an area under the receiver operating characteristic curve of 0.91.CONCLUSIONS: A machine learning model is capable of discerning MR on transthoracic echocardiography. This is an encouraging step toward machine learning-based diagnosis of valvular heart disease on pediatric echocardiography.
cardiac & cardiovascular systems