For the roughly 1-in-500 adults with hypertrophic cardiomyopathy, whether the heart's outflow tract is dangerously obstructed determines everything from medication choice to eligibility for septal reduction surgery. The catch: accurately measuring that obstruction has always required skilled Doppler echocardiography — a technically demanding step that creates bottlenecks in community cardiology settings and introduces significant operator variability. A model that bypasses Doppler entirely could democratize HCM triage at scale.
Published in Circulation: Cardiovascular Imaging, this study trained a cross-attention-based deep learning framework — built on the EchoPrime foundational video model — to classify left ventricular outflow tract gradients above 20 mmHg using only standard 2D transthoracic echocardiographic views. The derivation cohort comprised 1,833 patients from a U.S. tertiary center, with validation on a held-out internal set of 275 and a geographically distinct Korean external cohort of 46 patients. Single-view AI baselines performed modestly (AUC 0.47–0.70 externally), but the echo-pretrained EchoPrime single-view models improved substantially (AUC 0.75–0.83). The late-fusion multi-view model reached an AUC of 0.84 on the external cohort — a meaningful gain, though not yet at clinical deployment threshold.
This work sits at an important intersection: it demonstrates that temporal modeling from routine 2D cine loops can capture hemodynamic information previously thought to require spectral Doppler, likely by learning structural proxies such as systolic anterior motion of the mitral valve and septal geometry. The EchoPrime pretraining — trained on massive echocardiographic video datasets — appears to be the critical ingredient, outperforming generic vision models substantially. However, several limitations temper enthusiasm. The external cohort is small (N=46), limiting statistical confidence in generalizability. The binary threshold of 20 mmHg is a screening cut — clinical decisions often hinge on gradients above 50 mmHg. And HCM obstruction is dynamic, often requiring provocation maneuvers that static resting echo cannot replicate. This is an incremental but technically credible advance that sets the stage for larger, prospective validation.