Deep learning model predicts post-TAVR LVOTO in patients with severe aortic stenosis

Photo by Enchanted Tools / Unsplash

medRxiv Published April 10, 2026 Medically reviewed April 25, 2026 Study authors: Choi, J.-W.; Park, J.; Yoon, Y. E.; Kim, J.; Jeon, J.; Jang, Y.; Lee, S.-A.; Bak, M.; Choi, H.-M.; H… DOI ↗ By Dr. Amelia Tan, PhD · Internal Medicine & Chronic Disease

Key Takeaway

Consider the deep learning model as a potential predictor of post-TAVR LVOTO, though retrospective evidence limits causal claims.

A retrospective observational study examined 302 consecutive patients undergoing transcatheter aortic valve replacement (TAVR) for severe aortic stenosis. The primary exposure was the application of a pre-trained deep learning model to pre-TAVR transthoracic echocardiography, compared against conventional transthoracic echocardiography parameters and pre-TAVR left ventricular outflow tract (LVOTO) status. The primary outcome was post-TAVR LVOTO, defined as a peak pressure gradient of 30 mmHg or greater on follow-up imaging, assessed at a median of 47 days (IQR 37-63) after the procedure.

After multivariable adjustment, the deep learning index (DLi-LVOTO) was independently associated with the development of post-TAVR LVOTO. The adjusted odds ratio was 1.29 per 10-score increase in the DLi-LVOTO (95% CI 1.06-1.56; p=0.011). In the total cohort, 35 patients (11.6%) experienced post-TAVR LVOTO. Notably, the model retained independent predictive value even in patients who did not have pre-TAVR LVOTO, with an adjusted odds ratio of 1.56 per 10-score increase (95% CI 1.19-2.06; p=0.001).

The discriminative performance of the model was assessed using area under the receiver operating characteristic curve (AUROC). For the overall population, the AUROC was 0.78 (95% CI 0.72-0.85). In the subgroup of patients without pre-TAVR LVOTO, the AUROC was 0.84 (95% CI 0.77-0.91). Safety data, including adverse events or discontinuations, were not reported in this study.

Key limitations include the retrospective study design, which precludes causal inference. The findings suggest that the deep learning model captures hemodynamic features beyond conventional echocardiographic assessment. However, due to the observational nature of the evidence, these results should be interpreted as associations rather than proof of efficacy. Prospective studies are needed to confirm clinical utility and safety before routine implementation.

Study Details

Sample sizen = 32

EvidenceLevel 5

PublishedMar 2026

View Original Abstract ↓

AimsDynamic left ventricular outflow tract obstruction (LVOTO) is a hemodynamically significant complication following transcatheter aortic valve replacement (TAVR) that remains difficult to predict with conventional transthoracic echocardiography (TTE). We examined whether a deep learning (DL) model developed for LVOTO detection in hypertrophic cardiomyopathy (HCM) could predict post-TAVR LVOTO from pre-TAVR TTE in patients with severe aortic stenosis (AS). Methods and ResultsIn this retrospective study of 302 consecutive patients undergoing TAVR for severe AS, a pre-trained DL model was applied to pre-TAVR TTE to generate a patient-level DL index of LVOTO (DLi-LVOTO; range 0-100). Post-TAVR LVOTO was defined as a peak pressure gradient [≥]30 mmHg on follow-up TTE. Logistic regression and receiver operating characteristic analyses assessed the association and discriminative performance of DLi-LVOTO. Pre-TAVR LVOTO was present in 32 patients (10.6%) and post-TAVR LVOTO in 35 (11.6%). Follow-up TTE was performed at a median of 47 days (IQR 37-63) after TAVR, with the majority of TTE (216 of 302, 71.5%) performed within 2 months. DLi-LVOTO was significantly higher in patients with LVOTO at both pre- and post-TAVR TTE (all p<0.001). In multivariable analysis, DLi-LVOTO remained independently associated with post-TAVR LVOTO even after adjusting for conventional TTE parameters and pre-TAVR LVOTO (adjusted OR 1.29, 95% CI 1.06-1.56 per 10-score increase, p=0.011), with an AUROC of 0.78 (95% CI 0.72-0.85). Among patients without pre-TAVR LVOTO, DLi-LVOTO retained independent predictive value (adjusted OR 1.56, 95% CI 1.19-2.06, p=0.001; AUROC 0.84, 95% CI 0.77-0.91). ConclusionA DL model originally trained in HCM patients independently predicts post-TAVR LVOTO from pre-TAVR TTE, including in patients without pre-existing LVOTO, suggesting it captures hemodynamic features beyond conventional echocardiographic assessment.

Deep learning model predicts post-TAVR LVOTO in patients with severe aortic stenosis

Study Details

Evolocumab reduces cardiovascular events in diabetes without known atherosclerosis

Evolocumab cuts heart events in high-risk diabetes patients

Clinical research that matters. Delivered to your inbox.

Deep learning model predicts post-TAVR LVOTO in patients with severe aortic stenosis

More on severe aortic stenosis

Study Details

Evolocumab reduces cardiovascular events in diabetes without known atherosclerosis

Evolocumab cuts heart events in high-risk diabetes patients

Clinical research that matters. Delivered to your inbox.

Related in Cardiology

From Other Specialties