Left heart structure and function, as well as some clinical outcomes, can be accurately quantified by interpreting deep-learning echocardiography, a new study shows. These data were reported by Emily S. Lau, MD, MPH, of Massachusetts General Hospital, Boston and the Broad Institute of Harvard University and the Massachusetts Institute of Technology, Cambridge, Massachusetts, Paolo Di Achille, PhD, also of the Broad Institute, and colleagues in a manuscript published Monday online and in the Nov. 14 issue of the Journal of the American College of Cardiology. Noninvasive and widely available, echocardiography is used to understand cardiac structures and functions and diagnose a variety of cardiovascular(CV) disease conditions. However, interpretation of echocardiograms is time-consuming and requires an extensive amount of expertise. Previous studies suggest using deep-learning interpretation for a more efficient way to get results from echocardiograms. The investigators in this study developed the Dimensional Reconstruction of Imaging Data (DROID) model and studied how well it interprets standard measurements of left atrial (LA) and left ventricular (LV) structure and function, as well as incident CV outcomes. The DROID model is a validated three-dimensional convolutional neural-network model to classify and quantify LA dimension, LV wall thickness, chamber diameter and ejection fraction. A total of 64,026 echocardiograms (n=27,135 participants) were used in this retrospective, multicenter study, and were derived from electronic health records. A separate longitudinal primary care sample, along with an external health care system data set, were used for validation. Model-derived left heart measures with incident outcomes were evaluated using Cox models. Echocardiographic views (area under the receiver operating curve >0.97 for parasternal long axis, apical four-chamber and apical two-chamber views vs human expert annotation) and quantified standard left heart measures (R2 range=0.53 to 0.91 vs study report values) were distinguished with the deep-learning model. The DROID model performed similarly with the two external samples used for validation. Incident heart failure, atrial fibrillation, myocardial infarction and death were predicted by the deep-learning model by analyzing the left heart measures. A 1-standard deviation lower model-left ventricular ejection fraction was associated with a 17% risk of death (hazard ratio [HR]=1.17; 95% confidence interval [CI]=1.06-1.30) as well as a 43% greater risk of heart failure (HR=1.43; 95% CI=1.23-1.66). Other measures of the left heart were similar when derived from the model. Limitations of this study included potential bias in patient selection due to the retrospective nature of the echocardiographic videos used for training and validation of the deep-learning model and, thus, generalizability to other populations. In addition, the deep-learning model was trained with video selected by experts in the field at an academic medical center, leaving potential bias. Overall, the deep-learning echocardiography interpretation was able to predict some clinical cardiac outcomes and quantify standard measures of left heart structure and function. Clinical outcomes were associated with several differences in left heart structure and function. In an accompanying editorial, Márton Tokodi, MD, PhD, and Attila Kovács, MD, PhD, both of the Semmelweis University, Budapest, Hungary, provided some background, new insight and hopeful outlook for research in how deep learning may be used in the management of cardiovascular diseases. “[Artificial intelligence] AI-powered machines will undoubtedly play an increasing role in health care, not as a replacement but rather as a valuable supplement to health care professionals. Considering the high volume of studies ordered, echocardiography would significantly benefit from AI-powered solutions,” the editorialists wrote. Tokodi and Kovács pointed out many important aspects of this study, including validation strategies for the DROID model and the prediction process. The editorialists suggested a good next step would be conducting prospective, randomized trials to examine patient outcomes and look at costs. “Echocardiography, as shown by the success of the DROID models, is poised to reap substantial benefits from these advancements. As we move forward, we should embrace these AI-powered tools as invaluable partners, enhancing our ability to provide exceptional care to our patients,” the editorialists concluded. Sources: Lau ES, Di Achille P, Kopparapu K, et al. Deep Learning-Enabled Assessment of Left Heart Structure and Function Predicts Cardiovascular Outcomes. J Am Coll Cardiol. 2023;20:1936-1948. Tokodi M, Kovács A. A New Hope for Deep Learning-Based Echocardiogram Interpretation: The DROIDs You Were Looking For. J Am Coll Cardiol. 2023;20:1949-1952. Image Credit: ckybe – stock.adobe.com