A decision tree model combining deep learning-based strain and LVEF defined three risk profiles with 89.5% accuracy that predicted pre-severe AS systolic dysfunction and post-diagnosis mortality.
Cohort (n=602)
Does deep learning-based strain (DLS) predict future LVEF deterioration and mortality in patients with severe aortic stenosis?
Deep learning-based strain metrics combined with standard echocardiography can identify subclinical myocardial dysfunction up to 4 years before overt LVEF decline in patients with severe aortic stenosis.
Abstract Background Aortic stenosis (AS) imposes chronic pressure overload on the left ventricle (LV), leading to progressive myocardial remodeling and, eventually, systolic dysfunction. Identifying early markers of functional deterioration is essential to optimize monitoring and timing of intervention. Purpose To characterize the longitudinal evolution of LV systolic function in patients with AS, identify early predictors of LVEF decline, and assess their prognostic value — with a focus on deep learning-based strain (DLS) metrics. Methods We retrospectively studied 602 patients with severe AS (mean age 77 years, 54% female), all of whom had at least two prior echocardiograms (median 3 studies per patient over 4 years). At the time of severe AS diagnosis, patients were categorized as having low fraction (LF, LVEF 50%, n=124) or normal fraction (NF, LVEF ≥50%, n=478). Clinical characteristics, comorbidities, and echocardiographic parameters — including LV morphology and DLS metrics obtained with EchoNet-Dynamic — were compared. A decision tree model was developed to stratify risk profiles and evaluated for prognostic relevance. Results Patients in the LF group exhibited a pronounced decline in LVEF in the two years preceding AS diagnosis, while NF patients maintained stable systolic function (LVEF 58%). LVEF and DLS values were strong early predictors of future LVEF deterioration, detectable up to 4 years prior to the onset of severe AS. The decision tree model defined three echocardiographic profiles with 89.5% internal validation accuracy: high risk (LVEF 56%), intermediate risk (LVEF ≥56% and |DLS| 16%), and low risk (LVEF ≥56% and |DLS| ≥16%). These categories were independently associated with the likelihood of pre-severe AS systolic dysfunction and post-diagnosis mortality, regardless of whether aortic valve replacement was performed. Conclusion Subclinical myocardial dysfunction precedes overt LVEF reduction in patients with AS. Combining DLS with standard echocardiographic assessment enables earlier identification of at-risk individuals, potentially informing follow-up strategies and improving the timing of therapeutic intervention.
Lobo et al. (Thu,) conducted a cohort in Severe aortic stenosis (n=602). Deep learning-based strain (DLS) and LVEF assessment vs. Normal fraction (LVEF ≥50%) / Low risk profile was evaluated on Pre-severe AS systolic dysfunction and post-diagnosis mortality. A decision tree model combining deep learning-based strain and LVEF defined three risk profiles with 89.5% accuracy that predicted pre-severe AS systolic dysfunction and post-diagnosis mortality.