Deep-learning model achieved high diagnostic performance for detecting cardiac enlargement using electrocardiograms
1. Zhu and colleagues developed a deep learning model to detect atrial enlargement and ventricular hypertrophy or dilation using electrocardiograms (ECGs).
2. The model achieved greater average sensitivity than human interpretation, with reduced four-lead models retaining strong diagnostic performance.
Evidence Rating Level: 2 (Good)
Study Rundown: Cardiac hypertrophy, dilation, and enlargement are important causes of heart failure and sudden death. However, conventional diagnosis using echocardiography can be cost-prohibitive. ECG analysis using deep learning approaches has the potential to expand the identification of cardiac enlargement. Zhu and colleagues developed ECG-Cor-Net, a deep-learning model intended to improve detection using ECGs alone. Researchers assembled paired ECG and echocardiography records, using echocardiographic measurements as reference labels. The model classified ECGs as left or right ventricular hypertrophy or dilation, left or right atrial enlargement, or normal findings. Diagnostic performance was compared with interpretation by six physicians. The study found that the sensitivity increased from 0.270 to 0.586 after using the deep learning model. Models using four selected ECG leads instead of 12 leads retained comparable discrimination, although their ability to predict every label correctly was lower than that of the full model. This study demonstrates that artificial intelligence (AI) can be used to assist ECG screening for cardiac enlargement, but confirmatory echocardiography remains vital.
Click here to read the study in NEJM AI
Relevant Reading: AI-Enabled ECG Analysis Improves Diagnostic Accuracy and Reduces False STEMI Activations: A Multicenter U.S. Registry
In-Depth [retrospective cohort]: 80,007 ECGs from 65,927 adults were used for model development and held-out testing. The data were divided by patient identifier into training, validation, and testing sets in an 8:1:1 ratio. Diagnostic labels for the ECGs were supplied by echocardiographic measurements of chamber dimensions and wall thickness. Nine training runs were evaluated, with an additional test set of 473 ECGs from 469 patients being used to compare against the interpretations of six physicians. The outcomes included sensitivity, specificity, discrimination, and subset accuracy, defined as correctly predicting every label for an ECG. The model achieved an average sensitivity of 0.586 (standard deviation ± 0.053) versus 0.270 for physicians. Model sensitivities for right ventricular abnormalities, right atrial enlargement, left ventricular abnormalities, and left atrial enlargement were 0.440 (± 0.052), 0.417 (± 0.042), 0.565 (± 0.080), and 0.612 (± 0.064), respectively. These sensitivities were all higher than those of physicians. Four-lead models using I or II with aVR, V1, and V5 achieved subset accuracies of 0.551 and 0.534, respectively. This study was limited by retrospective sampling, a predominantly Chinese adult population, exclusion of depolarization abnormalities from test datasets, and limited external cases for some diagnoses. Nonetheless, this study demonstrated that deep learning models can be used to expand detection of cardiac enlargement, but prospective screening utility still needs to be determined prior to wider adoption.
Image: PD
©2026 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.
