Detection of Left Ventricular Outflow Obstruction from Standard B-Mode Echocardiogram Videos using Deep Learning

IntroductionHypertrophic cardiomyopathy (HCM) affects 20 million individuals globally, with increased risk of sudden death and heart failure. While cardiac myosin inhibitors show great promise as disease-specific treatment, current indications are for obstructive HCM. Obstruction is not always well characterized by echocardiography, and artificial intelligence (AI) might assist in the improving the underdiagnosis of left ventricular outflow tract (LVOT) obstruction. MethodsWe identified 2,693 patients with LVOT obstruction and 6,177 control patients matched by age, sex, and septal thickness. A deep learning model was trained on non-Doppler apical-4-chamber (A4C) B-mode echocardiographic videos to detect the presence of outflow obstruction identified later in the same study by spectral Doppler. Model performance was evaluated on held-out test sets from Cedars-Sinai Medical Center (CSMC) and from Stanford Healthcare (SHC). ResultsIn a held-out test set of 6,034 videos from CSMC, our model demonstrated strong performance in detecting LVOT obstruction with an AUC of 0.865, sensitivity of 77.8%, and specificity of 79.8%. Performance was consistent across various patient subgroups, including those with hyperdynamic LV function, pre-existing valvular disease, and small LV cavity size. The model demonstrated generalizable performance in the SHC cohort with an AUC of 0.866. ConclusionIn this study, we developed an AI model to detect LVOT obstruction from standard A4C echo videos, highlighting patients may benefit from more detailed cardiac workup for obstructive HCM.