Validation of an AI Echocardiography Model for Detecting LVOT Obstruction

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Circulation: Cardiovascular Imaging, https://doi.org/10.1161/CIRCIMAGING.126.019956 
Dr. Owen R. Crystal  et al.

Points

Mayo Clinic researchers developed an artificial intelligence model to identify left ventricular outflow tract obstruction using routine resting two-dimensional B-mode echocardiogram videos.
Approximately two-thirds of hypertrophic cardiomyopathy patients develop left ventricular outflow tract obstruction that complicates cardiac output and guides therapeutic interventions.
The model was trained on 1,833 patients, tested in 275 individuals, and externally validated in 46 patients at a hospital in South Korea.
Combining spatial and temporal data across three standard ultrasound views allowed the AI system to outperform expert echocardiographers reviewing non-Doppler imagery.
Automated B-mode image screening provides cost savings compared to universal Doppler testing and enables point-of-care obstruction screening on portable ultrasound units.

Summary

This study evaluated the diagnostic performance and generalizability of a deep learning artificial intelligence (AI) model designed to identify left ventricular outflow tract (LVOT) obstruction in hypertrophic cardiomyopathy (HCM) from resting B-mode echocardiographic videos without Doppler input. Approximately two-thirds (67%) of HCM patients develop LVOT obstruction, a key hemodynamic parameter guiding clinical decision-making. Led by Imon Banerjee at Mayo Clinic, the investigation sought to extract subtle spatial-temporal patterns from routine two-dimensional ultrasound videos across three standard views to detect elevated LVOT pressure gradients and inducible obstruction.

The development cohort comprised 1,833 patients from Mayo Clinic, with model validation conducted in an internal test subset of 275 patients and external validation performed in a geographically distinct cohort of 46 patients from South Korea. Combining spatial-temporal data across multiple standard echocardiographic views significantly improved predictive accuracy for LVOT gradient classification compared to single-view assessments. In sub-analyses evaluating non-Doppler image interpretation, the AI model outperformed two expert echocardiographers in detecting LVOT obstruction, demonstrating that hemodynamic gradients leave subtle structural signatures in two-dimensional B-mode cineloops.

External validation confirmed robust model performance across diverse patient populations and echocardiographic platforms despite baseline demographic and structural variations. Economic modeling demonstrated substantial cost savings for the AI-guided triage approach compared with universal Doppler screening protocols. The authors conclude that AI-based B-mode video analysis complements standard Doppler echocardiography, offering real-time decision support and extending non-invasive LVOT obstruction screening to point-of-care, portable, or resource-limited clinical settings.

Link to the article: https://www.ahajournals.org/doi/10.1161/CIRCIMAGING.126.019956

References

Crystal, O. R., Farina, J. M., Scalia, I. G., Ayoub, C., Etchegoyen, C. V., Chollet, L., Park, H. B., Kim, K. A., Arsanjani, R., Lester, S. J., & Banerjee, I. (2026). Beyond doppler: Scalable ai detection of lvot obstruction in hcm. Circulation: Cardiovascular Imaging, e019956. https://doi.org/10.1161/CIRCIMAGING.126.019956