A deep learning model was developed to detect severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV) using chest X-rays (CXRs). The model was trained on 71,589 CXRs from 24,689 patients, with labels derived from echocardiograms. The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.79 for SLVH, 0.80 for DLV, and 0.80 for the composite label (SLVH/DLV). It outperformed 15 board-certified radiologists in detecting these abnormalities, achieving a sensitivity of 71% compared to 66% for the consensus vote at a fixed specificity of 73%. The model was validated on an external dataset and showed similar performance. Saliency maps revealed that the model was sensitive to the cardiac silhouette and areas around the left ventricle. The study demonstrated that deep learning can accurately detect structural abnormalities in CXRs, potentially improving early diagnosis of heart failure. The model's performance was consistent across different populations and subgroups, including those with pacemakers, heart transplants, and lung transplants. The model's results suggest that CXRs can be used effectively for early detection of heart failure, even in asymptomatic patients. The study highlights the potential of deep learning in cardiology for early diagnosis and improved patient outcomes. The model and data are publicly available for further research.A deep learning model was developed to detect severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV) using chest X-rays (CXRs). The model was trained on 71,589 CXRs from 24,689 patients, with labels derived from echocardiograms. The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.79 for SLVH, 0.80 for DLV, and 0.80 for the composite label (SLVH/DLV). It outperformed 15 board-certified radiologists in detecting these abnormalities, achieving a sensitivity of 71% compared to 66% for the consensus vote at a fixed specificity of 73%. The model was validated on an external dataset and showed similar performance. Saliency maps revealed that the model was sensitive to the cardiac silhouette and areas around the left ventricle. The study demonstrated that deep learning can accurately detect structural abnormalities in CXRs, potentially improving early diagnosis of heart failure. The model's performance was consistent across different populations and subgroups, including those with pacemakers, heart transplants, and lung transplants. The model's results suggest that CXRs can be used effectively for early detection of heart failure, even in asymptomatic patients. The study highlights the potential of deep learning in cardiology for early diagnosis and improved patient outcomes. The model and data are publicly available for further research.