An artificial intelligence (AI) model was trained in echocardiography video for the automatic detection of hypertrophic cardiomyopathy (HCM). With validation, the model was able to differentiate HCM from phenocopies and detect obstruction in HCM gradients. 

The study, presented as late-breaking research during the American Society of Echocardiography (ASE) 2025 Scientific Sessions, showed the potential of AI to improve diagnostic accuracy of HCM.

"The deep learning model can detect hypertrophic cardiomyopathy, and that's all validated internally and externally with excellent results," said co-study author Fawaz Alenezi, MD, Associate Professor of Medicine, Duke Heart Center, Durham, North Carolina, when presenting the findings during the ASE 2025 meeting. 

HCM is characterized by unexplained left ventricular hypertrophy, measuring wall thickness in end diastole of at least 15 mm, without the presence of another disease that could explain why the wall thickness is so great. There are a wide range of clinical presentations and phenotypes for HCM and it has a diverse natural history. 

During the presentation Dr. Alenezi highlighted the challenges of diagnosing HCM and stressed the serious issue that 80% of patients with HCM are undiagnosed. "Sometimes it takes up to four clinical visits after the echo and up to 5 years to really diagnose a case of hypertrophic cardiomyopathy," he said. 

Transthoracic echocardiography (TTE) is most common noninvasive imaging modality used for the evaluation and diagnosis of HCM based on the assessment of septal wall thickness, left ventricular cavity dimensions, systolic function, and dynamic left ventricular outflow tract gradients. Further systolic anterior motion of the mitral valve, mitral-septal contact, and hyperdynamic left ventricular ejection fraction also help in diagnosing HCM.

Early forms of HCM can often be missed or diagnoses can be cofounded by cardiac diseases with similar phenotypes, including hypertensive heart disease, aortic stenosis, and cardiac amyloidosis. Researchers believe that use of trained deep learning approaches is promising for improving the accuracy of HCM detection, as AI could be taught to recognize the complex spatiotemporal patterns of the disease.

The study authors developed a deep learning model for the automated detection of HCM through routine echocardiographic video loops.

The AI model, which was processed with FDA- and CE-reviewed US2.AI software from Singapore, was developed for automated view classification and cardiac cycle segmentation. It was trained on a large dataset of a retrospective cohort of 1,472 adult patients from Duke University Medical Center who had expert-adjudicated HCM, as well as 7,292 control participants without HCM. Validation was then completed both internally and externally. The researchers assessed the ability of the model to distinguish between HCM and other forms of left ventricular hypertrophy.

The model learned from raw video data and provided a continuous probability score between 0 and 1 to portray the likelihood of the patient having HCM according to a
combination of view-specific predication and ensemble-level aggregation data.

To classify and detect gradients of HCM, the model was processed with US2.AI for automated Doppler view classification and Valsalva detection. The model generated continuous wave Doppler velocity time integral segmentation masks to calculate peak gradients. The researchers used 30 mmHg or more as the cutoff per AHA/ACC 2024 guidelines. These AI-derived gradients were compared with ground-truth resting and provocation values from the clinical reports of 712 patients with HCM from Duke University Medical Center. 

From the overall Duke dataset, 818 cases of HCM and 3,470 controls were withheld to be used in an internal validation set. The researchers assessed the model’s diagnostic accuracy with the validation set and generalizability with a Japanese external validation cohort of 94 patients with clinically confirmed HCM and 351 patients with other forms of left ventricular hypertrophy.

Patients in the Duke training cohort had an average of 59 years, and 45% were male. The average septal wall thickness was 12.5 mm.

In the patients with HCM, 35% had a resting left ventricular outflow track gradient of more than 30 mmHg and 21.3% had a provoked gradient with Valsalva of more than 30 mmHg, indicating dynamic obstruction. Controls participants had similar characteristics without evidence of HCM.

In the training cohort, the AI model achieved 99.6% accuracy with a 98.9% (95% confidence interval [CI] = 98.2%–99.6%) sensitivity and 99.7% (95% CI = 99.6%–99.9%) specificity. The positive predictive value was 65.7% (95% CI = 50.3%–77.7%).

With the Duke validation or internal testing cohort, the AI model showed an accuracy of 93.4%, sensitivity of 80.2%, and specificity of 94.9%. The AI model achieved an area under the curve (AUC) of 0.95 for being able to distinguish between HCM and controls among participants. Specifically, when differentiating between HCM and cardiac amyloidosis, the AUC was 0.95, with a sensitivity of 80% and a specificity of 93%; for HCM vs aortic stenosis, the AUC was 0.83, the sensitivity rate was 80%, and the specificity rate was 69%; and for HCM versus hypertensive heart disease, the AUC was 0.85, a sensitivity of 80%, and a specificity of 75%. The positive predictive value was 7.3%.

The model also showed good performance in subgroup analyses with AUCs of 0.94 for patients with interventricular septal thickness or left ventricular posterior wall thickness at end diastole of at least 13 mm, and 0.93 for patients with interventricular septal thickness and left ventricular posterior wall thickness at end diastole of under 13 mm.

This indicated, according to the study authors, that the model was capable of detecting HCM even with more mild morphological presentations.

In the independent external Japanese validation cohort, accuracy measured 82.5% with a sensitivity of 79.8% and a specificity of 83.2%. The AUC was 0.9, which supported the generalizability of the model across international and ethnically diverse groups. The positive predictive value was 2.3%.

When detecting gradients of HCM in this cohort of the external validation dataset—for the detection of obstruction—the sensitivity was 91%, the specificity was 82%, and the AUROC for accuracy was 0.91. 

For the continuous wave gradient at rest, the correlation was 0.77 with a yield of 81.5% compared with 0.8 and 78.8% at Valsalva or when provoked. The overall external validation showed a correlation of 0.77 and yield of 80.4%. 

The researchers proved that the model was not led predominantly by one measurement or imaging view over another in its decision process but rather by an integrated multi-view and temporal information approach.

No significant differences were observed in the model’s performance across different echocardiographic vendors, image acquisition protocols, or patient gender. Any variations in performance observed across age subgroups was found to be statistically insignificant.

Alenezi and the study authors concluded that their findings showed the feasibility of using a video-based AI model for automated detection of HCM from echocardiography imaging. He stressed that the model needs further study and validation. Going forward, they plan to prospectively validate the model with deployment in real-time echocardiographic workflows and try to integrate the model with other data for more comprehensive phenotyping and longitudinal monitoring.