Overview

Hypertrophic cardiomyopathy is the most common inherited cardiomyopathy and a leading cause of sudden cardiac death in the young. It is now treatable in the era of cardiac myosin inhibitors, yet remains substantially underdiagnosed. Published expert cardiologist sensitivity for echocardiographic HCM detection averages only 50.7%. Prior AI approaches have required manual view selection, annotated measurements, or report fusion, and none has unified detection, phenocopy discrimination, and outflow obstruction assessment in a single deployable pipeline. This late-breaking presentation described a fully automated system that addresses all three tasks from routine transthoracic echocardiography with no manual input.

 

Study Design

Two video-based Multiscale Vision Transformer v2-small models were developed.

Model 1 for HCM detection was trained on 9,133 patients from Duke University Medical Center and the National Cerebral and Cardiovascular Center, Japan, including HCM cases, LVH phenocopies (cardiac amyloidosis, aortic stenosis, hypertensive heart disease), and matched controls. Probability thresholds were prespecified on internal validation and frozen before external testing. External validation used 791 independent patients from Rutgers Robert Wood Johnson University Hospital.

 

Model 2 for continuous-wave Doppler LVOT gradient estimation was trained on 2,228 sonographer-annotated VTIs from 407 HCM patients at the Brigham Cardiovascular Imaging Core Laboratory, with external validation in 716 independent HCM patients from Duke.

 

Key Results

 

HCM Detection (Model 1):

  • Internal test set: AUC 0.97, accuracy 93.8%, sensitivity 92.6%, specificity 94.0%

 

  • External validation at Rutgers: AUC 0.96 against general controls (sensitivity 87.4%, specificity 95.2%) and AUC 0.93 against LVH phenocopies (sensitivity 87.4%, specificity 89.3%)

 

  • Performance preserved across obstructive and non-obstructive phenotypes and in patients without overt hypertrophy
  • Model discriminated HCM from cardiac amyloidosis (AUC 0.97), aortic stenosis (AUC 0.97), and hypertensive heart disease (AUC 0.85)

 

LVOT Gradient Estimation (Model 2):

  • External validation cohort of 716 HCM patients: ICC 0.77 (95% CI 0.73 to 0.81), mean absolute error 17.6 mmHg
  • Sensitivity 90%, specificity 81%, AUROC 0.88 for identifying obstructive physiology at the guideline threshold of 30 mmHg or above

 

Why This Matters

HCM is underdiagnosed at rates that carry significant clinical consequences, particularly in the current era of effective cardiac myosin inhibitor therapy. A system that detects HCM from routine echocardiography with sensitivity exceeding expert cardiologists, distinguishes it from clinically relevant mimics, and quantifies LVOT obstruction without manual input represents a meaningful advance for population-scale screening and treatment eligibility assessment. The preservation of performance in patients without overt hypertrophy confirms that the model captures features beyond wall thickness thresholds, aligning with the biological complexity of HCM.

 

Conclusion

A fully automated end-to-end deep learning system detects HCM, distinguishes it from clinically relevant LVH phenocopies, and quantifies LVOT obstruction from routine echocardiography with no manual selection or annotation, exceeding published expert benchmarks and positioned to address underdiagnosis and standardise functional assessment for treatment selection at population scale.

 

Alenezi, F. (2026, August 28-31). Deep Learning Model for Detection, Phenotyping, and Left Ventricular Gradient Estimation in Hypertrophic Cardiomyopathy Using Echocardiography: Development and Multinational External Validation [Late-breaking presentation]. ESC Congress 2026, Munich, Germany.